Control of Connected and Automated Vehicles: State of the Art and Future Challenges
CControl of Connected and Automated Vehicles: State of the Art and Future Challenges
Jacopo Guanetti a, ∗ , Yeojun Kim a , Francesco Borrelli a a Department of Mechanical Engineering, University of California, Berkeley, CA, 94720 USA.
Abstract
Autonomous driving technology pledges safety, convenience, and energy e ffi ciency. Challenges include the unknown intentions ofother road users: communication between vehicles and with the road infrastructure is a possible approach to enhance awareness andenable cooperation. Connected and automated vehicles (CAVs) have the potential to disrupt mobility, extending what is possiblewith driving automation and connectivity alone. Applications include real-time control and planning with increased awareness,routing with micro-scale tra ffi c information, coordinated platooning using tra ffi c signals information, eco-mobility on demandwith guaranteed parking. This paper introduces a control and planning architecture for CAVs, and surveys the state of the art oneach functional block therein; the main focus is on techniques to improve energy e ffi ciency. We provide an overview of existingalgorithms and their mutual interactions, we present promising optimization-based approaches to CAVs control and identify futurechallenges. Contents1 Introduction 12 System Components 2 ∗ Corresponding author
Email addresses: [email protected] (Jacopo Guanetti), [email protected] (Yeojun Kim), [email protected] (Francesco Borrelli)
1. Introduction
Autonomous driving has been the object of great research ef-forts in the last decades. Human errors are a prominent causeof road accidents and fatalities. Road congestion causes in-e ffi ciency in daily commutes and other aspects of road trans-portation. A transportation system that is less reliant on humandrivers allows all the passengers to better use their travelingtime and is associated with fewer road accidents.While the idea has been around for almost a century, it was inthe 1980s that the technological advances in sensing and com-puting made it realistic. Much of the early research on auto-mated driving was in the field of automated highway systems.The California PATH program, started in 1986, demonstratedautomated driving on four vehicles on the I-15 in San Diegoin 1994 [1, 2]. Other early successes were the PROMETHEUSproject [3, 4] and the CMU NAVLAB [5], that demonstrated thecapability of driving for hundreds of miles with minimal humanintervention. More recently, various research groups have com-mitted to demonstrations of autonomous driving in a variety ofscenarios, including the DARPA Grand Challenge in 2004 [6],the DARPA Urban Challenge in 2007 [7], the Intelligent Vehi-cle Future Challenge [8], the Hyundai Autonomous Challengein 2010 [9], the VisLab Intercontinental Autonomous Challengein 2010 [10], the Public Road Urban Driverless Car Test in 2013[11], and the autonomous drive Bertha-Benz historic route [12].With a substantial body of knowledge and continuousimprovements in perception technologies and computationalpower, autonomous driving features are being slowly intro-duced in everyday life. While all major brands have introducedadvanced driving assistance systems, such as adaptive cruisecontrol and automatic emergency braking, massive research ef-forts are being put into self-driving cars. The list of players in-cludes manufacturers like Tesla [13], Ford [14], and GM [15],suppliers like Bosch [16] and Delphi [17], and tech corpora- Preprint submitted to Annual Reviews in Control April 12, 2018 a r X i v : . [ c s . S Y ] A p r ions like Google [18] and Uber [19]. The SAE standard J3016[20] has classified six levels of driving automation.Vehicle connectivity has also been maturing in the pastdecades. Connectivity enables many convenience features andservices, including emergency calls, toll payment, and infotain-ment. Connectivity has also emerged as a technology to im-prove safety, performance, and enable vehicle cooperation: inthe aforementioned California PATH program [1, 2], the con-cept of platoon (a group of vehicles traveling at small spacing)was demonstrated as a way to increase the throughput of au-tomated highways. Vehicle-to-Vehicle (V2V) communicationhad to purpose of coordinating multi-vehicle maneuvers, and -once a formation was established - to exchange vehicle states,enabling short headway time between vehicles. Other demon-strations of cooperative driving took place in the Demo 2000program in Japan in 2000 [21], in the Grand Cooperative Driv-ing Challenge in 2011 and 2016 in the Netherlands [22, 23], theSARTRE program [24], and the Energy-ITS program started in2008 in Japan [25].Internet connectivity is now present in several vehicles, andthe Dedicated Short Range Communication (DSRC) technol-ogy has been adopted for V2V and Vehicle-to-Infrastructure(V2I) applications by manufacturers like Cadillac [26] andAudi [27], and transposed by the SAE in the J2735 standard[28]. Developments in 5G technologies might also support V2Iand even V2V communication in the future.Connected and Automated Vehicles (CAVs) have the poten-tial to extend what is possible with driving automation and ve-hicle connectivity alone. Connectivity has the potential to dra-matically improve environment awareness, and thus safety, ofautonomous vehicles, in spite of limitations of perception sys-tems. Automation can make full use of connectivity, especiallyfast V2V communication (10 Hz or more). CAVs enable a vari-ety of applications in intelligent transportation systems, includ-ing tra ffi c control, cooperative driving, improved safety, and en-ergy e ffi cient driving [29], although the latter may be partiallyharmed by the increased sensing, computing and communica-tion equipment of CAVs [30].This survey is focused on a control and planning architecturefor CAVs, and particularly on approaches for the improvementof energy e ffi ciency. We review the state of the art for the mostrelevant functional blocks of this architecture, including real-time controls, real-time motion planning, eco-driving, multi-vehicle coordination, and routing. To limit the scope of thissurvey, we focus on vehicle controls rather than tra ffi c control,although, in some multi-vehicle applications, the di ff erence isblurred. For the same reason, we dismiss the aspects of per-ception and environment prediction, both crucial parts of anyautonomous driving control architecture.The remainder of this paper is structured as follows. In Sec-tion 2 we describe the main components of a typical CAV sys-tem. In Section 3 we describe the control and planning archi-tecture that is the scope of the paper. In Section 4 we survey thestate of the art for real-time control and planning algorithms,that are generally implemented on-board. In Section 5 we sur-vey the literature on longer-term planning and routing algo-rithms, that are generally implemented remotely. Some con- cluding remarks end the paper.
2. System Components
As shown in Figure 1, the successful deployment of CAVsin an intelligent transportation system depends both on the on-board instrumentation and on the surrounding environment, i.e.the road infrastructure (including signalized intersections, rampmeters, road signs) and the other road users (including otherCAVs, non-cooperative vehicles, cyclists, pedestrians). In thissection, we give a brief overview of CAVs and the agents theyusually interact with. While not strictly focused on algorithmicaspects, this short digression helps to evaluate the significanceand technical soundness of the algorithms discussed later.
We define as Connected and Automated Vehicle (CAV) a ve-hicle that is capable of automated driving and connectivity withother vehicles or road users, the road infrastructure, and thecloud. Thus, CAVs are distinguished by driving automation andconnectivity; powertrain control is not a peculiarity of CAVs,but it is important for applications oriented to energy e ffi ciencyimprovement. Vehicles Automation
Autonomous driving includes the control of vehicle motionin both the longitudinal and lateral direction. This in turn re-quires an interface to the powertrain and the steering system.GPS can provide positioning with an accuracy that variesfrom meters to centimeters, depending on the specific tech-nology and the environmental conditions. GPS and prior roadmaps enable navigation and localization of the vehicle with re-spect to static elements. To enable driving in dynamic environ-ments, autonomous driving systems include an array of sensorsfor environment perception, including lidar, radar, cameras, andultrasonic sensors. A high resolution perception system, com-bined with high resolution environment maps, can also providecentimeter-level localization; this is appealing e.g. in urban ar-eas, where GPS accuracy is low.
Vehicles Connectivity
Communication and connectivity are enabling technologiesfor intelligent transportation systems. Perception makes self-driving vehicles aware of their surroundings, similar to sensesfor human drivers. Multi-vehicle cooperation, awareness of ob-stacles outside the line of sight, and forecasts require communi-cation. Today’s core technologies are DSRC and cellular com-munication (4G and 5G).DSRC is a wireless communication technology, mostly con-ceived for active safety. The US Department of Transportation[31], the SAE [28], ETSI [32], and several private companies[33, 34, 35, 36, 26, 27] have used DSRC to develop standardsand products. Applications include safety warnings (forwardcollisions, blind spots, emergency vehicles, road works), inter-section assistance and safety, tra ffi c conditions, payment of tolls2 igure 1: Cartoon depicting a variety of intelligent transportation system on highway, arterial and urban roads enabled by connected and automated vehicles (CAVs).Each number refers to a CAV application discussed next. Communication with other vehicles enables (1) augmented awareness, (2) platooning, and (3) cooperativemaneuvers. Communication with the infrastructure enables (4) enhanced approach and departure to signalized intersections. Cloud connectivity enables access todatabases, forecasts, and remote computations. On-board perception, localization and maps are fundamental to navigate in known and unknown environments, thatcan include non-connected vehicles, cyclists, pedestrians. In (5) roadway sensors generate signal phase and timing (SPaT) and vehicle occupancy and speed (VOS)data, that can be stored in the cloud. Other applications include coordination of grid charging, parking, road works (6). (Created on https://icograms . com ). and parking. The DSRC technology may also be used to im-prove GPS accuracy [37] and for geo-fencing.Internet connectivity via cellular communication enables ac-cess to cloud-based data and services. Due to its low latency,5G may also compete with DSRC for V2V and V2I communi-cations.Advantages of the DSRC technology include security, low la-tency, interoperability, and resilience to extreme weather condi-tions; on the flip side, it requires dedicated hardware. 5G o ff ersboth access to multimedia and cloud services (that are highlyvalued by customers), and cooperation with other vehicles andinfrastructure. The solution in the future may be a combinationof both technologies [38]. Vehicle Powertrains
By controlling vehicle motion with increased awareness,CAVs can inherently improve energy usage. Additionally, pow-ertrain control systems can benefit from the forecasts that maps,perception, and communication make available.The majority of today’s powertrains are based on internalcombustion engines; more advanced powertrains (sometimescalled micro- and mild-hybrids) can include start / stop systems,engine coasting systems, and some energy regeneration [39].Hybrid electric vehicles (HEVs) have high voltage, mediumcapacity battery packs, an electric motor, and an internal com-bustion engine, allowing pure electric driving, pure thermaldriving, and hybrid driving. In plug-in HEVs, the battery can berecharged from the grid and the battery pack is typically largerthan in HEVs, allowing to drive on electricity for significantdistances.The appeal of purely electric vehicles is due to the absenceof local emissions, low price of electric energy, good dynamicperformance and low noise. In any powertrain, auxiliary loads (e.g. air conditioning,lights, infotainment) can significantly a ff ect the overall en-ergy consumption. For systems that are not safety-critical, thelevel of performance may be temporarily decreased to limit thepower consumption. The roads on which CAVs are operated include complex sys-tems for tra ffi c monitoring and control. Static and dynamicmaps, databases, and remote computations can be accessed byCAVs through the cloud. Highway infrastructure
Modern highways are instrumented with systems for vehi-cle detection, to monitor and potentially control the tra ffi c flow.A variety of detection technologies are employed, includingin-roadway sensors (loop detectors, magnetic detectors, mag-netometers) and over-roadway sensors (cameras, radars, ultra-sonic, infrared, and acoustic sensors) [40]. The uses of sensorsin highways include data collection (vehicle occupancy, speed,type) for monitoring and planning of road use [41], and activetra ffi c control via ramp metering. Urban infrastructure
A typical instrumented intersection (see e.g. [42]) featuresin-roadway sensors, like loop detectors or magnetic sensors,that detect the presence of vehicles at a stop bar. Additionalsensors at advance locations and in departure lanes can also beused to estimate the vehicles speed and turn movements [43].The signal phase and timing (referred to as SPaT and describingthe current light color and the remaining until the next changeof color) can be retrieved directly from the controller, or indi-rectly via image analysis.3he uses of these data include the analysis of intersectionsperformance, tuning of controllers, feedback to adaptive con-trollers, and broadcasting to vehicles for coordination and traf-fic flow improvement [44]. Controllers can implement a fixedcycle, change green times depending on immediate tra ffi c con-ditions, or implement more advanced control strategies adapt-ing to congestion level; pedestrians can be part of the cycleor make requests with buttons. Metrics for intersection per-formance include volume-to-capacity ratios, fraction of arrivalsin green, red-light violations, queue delays [45, 46]. These datamay be stored locally and collected by operators, or stored re-motely on the cloud.Intersections can be instrumented to broadcast messages tonearby vehicles using DSRC. The SAE standard J2735 [28] in-cludes messages for signal phase and timing and intersectiongeometry. Notice that the timing part is deterministic only ifthe controller has a fixed cycle; otherwise, the timing is inher-ently uncertain, because of the stochastic nature of tra ffi c.Beyond intersections and tra ffi c signals, urban infrastructureincludes fuel stations, charging stations, and parking infrastruc-ture. A connected charging and parking infrastructure enablesbetter routing of vehicles, and more e ff ective pricing schemes.Vehicle charging has a significant e ff ect on grid balancing andsmart grids [47, 48, 49]. Automated parking systems enablebetter exploitation of urban surfaces. Cloud infrastructure
Cloud services supply to CAVs static and dynamic roadmaps, historical databases, and remote computational power.Access to the cloud is enabled by cellular connections.Modern road and tra ffi c map services (see e.g. [50, 51, 52])provide information that goes beyond the maps for navigation,including: • static information, such as road grade, road curvature, lo-cation of intersections, lane maps, speed limits, location offuel and charging stations, intersection average delays; • dynamic information, such as tra ffi c speed, availability andprice of fuel and charging stations, intersection delays,tra ffi c congestion, road works, weather conditions.Historical data can be relevant for planning problems, likevehicle routing and reference velocity generation. Examples in-clude tra ffi c congestion on highways [41], and signal phase andtiming data [53, 54]; in both cases, historical data give deeperinsight on tra ffi c patterns, that is generally not found in maps.CAVs can perform computations remotely, using cloud ser-vices and partially alleviating the on-board computational re-quirements. Computations moved to the cloud may include(dynamic) routing and long-term trajectory optimization. Multi-vehicle cooperation happens among two or moreCAVs. The aforementioned SAE J2735 standard [28] is mostlyoriented to awareness and safety applications. Advanced ve-hicle cooperation, including multi-vehicle formations, require
Remote Planning and RoutingEco-RoutingEco-Drivingand CoordinationBattery ChargePlanning MapsHistorical DataRemoteOn-boardReal-Time Control and PlanningMotionPlanningMotionControlPowertrainControl P o w e r t r a i n I n t r e f ace S t ee r i ng I n t r e f ace Sensor Fusion(Map-based) LocalizationPerceptionEnvironment Prediction U I G PS / I N S D S RC R a d a r L i d a r C a m e r a Figure 2: Architecture for Connected and Automated Vehicles (CAVs) deploy-ment. more complex protocols [55], for which a standard has not yetbeen established. An appealing niche for cooperative vehiclesis freight transportation; heavy duty vehicles capable of auto-mated driving and V2V communication can form platoons anddrive at small inter-vehicular distance, thereby reducing theirair drag resistance and fuel consumption (see e.g. [56]).
Non-cooperative vehicles, cyclists and pedestrians
When interacting with non-cooperative road users, CAVs donot di ff er substantially from other self-driving vehicles. Inthis case, awareness of the surroundings relies on the percep-tion system. This includes non-cooperative vehicles, cyclists,pedestrians, and any other road user. Recent research and tech-nologies are oriented to some level of cooperation with cyclistsand pedestrians, enabling safety communications between ve-hicle and smartphones [33, 57].
3. Connected and Automated Vehicle Control Architecture
Figure 2 shows a control architecture for Connected and Au-tomated Vehicles (CAVs) focused on safe and energy e ffi cientoperation. The architecture includes on-board and remote func-tional blocks. The functional blocks that reside on-board are safety-critical,and need to be executed in real-time. The real-time layer inter-faces to the vehicle actuators, collects measurements from on-board sensors, and performs all the real-time computations thatmake a CAV reliable and robust to unpredicted events. These4omputations include the control and planning algorithms thatare shortly described next, and detailed in Section 4.
Powertrain control
Powertrain control depends on the powertrain type and mayinclude engine control, electric motor control, gear shiftingcontrol. Powertrain controls satisfy in real-time the power re-quired to move the vehicle, and a ff ect the so-called “tank-to-wheel” energy conversion [39, 58]. Reactive controls selectthe powertrain operating points based on the current power de-mand. Energy e ffi ciency can be improved when forecasts areavailable, both for the short-term (speed and torque profilesfrom longitudinal control) and long-term (from the cloud layer). Motion control
The motion control block regulates the longitudinal and lat-eral motion of the vehicle, and is interfaced to the powertraincontrols and the steering system. The desired vehicle motionis generally specified at a higher hierarchical control level, andthe motion control ensures that the reference behavior is exe-cuted in closed loop. Motion control a ff ects safety and the so-called “wheel-to-distance” energy conversion [39, 58]. Whenforecasts of tra ffi c, signals, and trajectories of other vehiclesare available, safety and performance can be significantly im-proved. Motion planning
The real-time planning block includes maneuver planning(e.g. decision to stay in a lane or change), path planning, andtrajectory planning. These blocks also depend on the drivingcontext, and their boundaries are quite blurred [59].
The remote layer in Figure 2 enables access to external datasources, and performs longer term computations, that mostlya ff ect performance and are not real-time critical. These com-putations include the planning and routing algorithms that areshortly described next, and discussed into details in Section 5,following a bottom-up order. Battery charge planning
If the CAV is an electric, hybrid, or plug-in hybrid electric ve-hicle, a long-term planning of the battery charge trajectory canprevent suboptimal utilization of the energy stored on-board. Inan electric vehicle, this algorithm can simply predict the driv-ing range using route information; if the range allowed by thecurrent battery charge is exceeded, the algorithm may alert theuser, request to re-plan the route, or plan a stop in a chargingstation. In hybrid and plug-in hybrid electric vehicles, an inter-nal combustion engine is available; the route information can beused to optimize the allocation of fuel power and battery poweralong the trip.
Eco-driving and coordination
The eco-driving and coordination block takes route informa-tion and computes a reference velocity trajectory for the on-board algorithms. The value of this block is in the use of long-term forecasts (like road grade and tra ffi c congestion) and in theaccounting for constraints like trip time and maximum velocity.Some constraints depend on the driving context: for instance,passing a signalized intersection during a green phase; histori-cal data may help to improve performance.In these driving scenarios, the ego-CAV can cooperate withother CAVs. An example of multi-vehicle coordination is pla-tooning , in which a group of vehicles travel on a certain roadsegment at reduced distance gaps [2, 60]. The objective can beto maximize the usage of road surface (and hence throughput)or to reduce the aerodynamic drag. In the multi-vehicle case,the eco-driving block uses the same information, but the prob-lem is generally more complex. Eco-routing
The eco-routing block determines the most energy-e ffi cientroute, given user requirements and road maps (e.g. road grade,tra ffi c speed, intersection delays, fuel or charging stations).This block outputs the optimal route, i.e. a set of waypointsalong with the intersection locations, speed limits, road grade. The real-time planning and control blocks require feedbackfrom the vehicle, its position and velocity relative to the sur-rounding environment, and predictions of moving obstacles[61]. A CAV may be equipped with a GPS unit for localiza-tion, cameras, radars and lidars for perception, and a DSRCunit for V2V and V2I communication. These data are processedand fused to estimate the position and velocity of the CAV andthe surrounding objects, both static and moving. To cope withagents like pedestrians, cyclists and non-connected vehicles, analgorithm predicts the future trajectories of moving obstacles.Perception, localization, and environment prediction are ex-tremely important for self-driving vehicles and CAVs. The in-terest of the academic and industrial research communities onthese topics is very high, and has produced a vast literature.To limit the scope of this survey, we only focus on the real-time control and planning layer and on the remote planning androuting layer . In the next two sections, we will analyze the functionalblocks in the real-time control and planning layer (in Section 4)and the remote planning and routing layer (in Section 5), fol-lowing a bottom-up approach. The actual inputs and outputsof each block will be more precisely specified, improving theunderstanding of the overall architecture. For each block, ourmain goal is to survey the existing literature. To limit the scopeof this survey, we mostly focus on optimization-based methodsand energy e ffi ciency, and we point to more focused surveys onspecific topics. Contextually, we highlight the challenges andopportunities enabled by CAVs. Opportunities are often related5ehicle type Fuel Electric (Plug-in)powered HybridGear shifting (cid:88) (cid:88) (cid:88) Engine on / o ff (cid:88) (cid:88) Energy management (cid:88)
Table 1: The powertrain control problems surveyed in this paper, and theirapplicability to the most common powertrains. to automated and cooperative driving, improved environmentforecasts, and connectivity for data and remote computations.Driver safety, performance improvement, and real-time opera-tion are identified as the main technical challenges; real-timeoperation includes the coordination between on-board and re-mote layers. Where pertinent, we illustrate selected approacheswith more detailed examples.
4. On-board real-time control and planning
In this section, we review the existing literature for each ofthe three functional blocks in the real-time control and plan-ning layer of Figure 2: powertrain control , motion control , mo-tion planning . Performance metrics include vehicle energy con-sumption, passenger comfort, and - at a broader level - roadthroughput. The main safety requirement is to avoid collisionswith other road users. This separation in blocks gives struc-ture and facilitates the review; nonetheless, the boundaries aresometimes blurred. Several of the works that we discuss inte-grate, at least partially, two or more of these blocks. Powertrain control has a broad meaning and includes manycomponents and subproblems, such as transmissions, internalcombustion engines, electric motors, starters and generators. Atlarge, powertrain controls address power generation for vehiclemotion and auxiliary loads. The literature on the topic is ex-tremely vast; in this section, we focus on three powertrain con-trol problems in which connectivity and driving automation areor can be leveraged.
Literature review Di ff erent powertrain architectures allow more or less flexi-bility in the realization of the power demand for vehicle motionand auxiliary loads. In this paper, we focus on fuel-powered ve-hicles, electric vehicles, hybrid and plug-in hybrid vehicles (seeFigure 3). We survey gear-shifting control, engine on / o ff con-trol, and energy management; Table 1 maps these three prob-lems to the di ff erent powertrain configurations. Gear-shiftingand engine on / o ff are self explanatory. By energy (or power)management in hybrid vehicles, we refer to the problem of al-locating the power demand to the internal combustion engineand the electric motor.In the three problems listed above, the goal is to minimize acost function of the form J = (cid:90) T (cid:16) γ f P f ( t ) + γ q P q ( t ) (cid:17) dt , (1) WTE (a) Internal combustion engine vehicle. WTMB (b) Electric vehicle.
WTMCE B (c) Pre-transmission or single-shaft parallel hybrid electric vehicle.Figure 3: Common powertrain topologies. Thin lines: electrical connections.Thick lines: mechanical connections. W: longitudinal dynamics. T: transmis-sion. E: internal combustion engine. M: electric motor. B: high-voltage battery.C: clutch. Further powertrain topologies, including series and combined hy-brids, are presented in [39]. where T is the duration of the driving schedule, P f is the powerextracted from the fuel, P q is the battery internal power, γ f and γ q are their weights. It is easy to determine the optimal policyfor a fixed profile of the power demand; for instance, the opti-mal gear shifting policy during a standard driving cycle may becomputed by dynamic programming. In real-time operation thepower demand is not known in advance, but the optimal pol-icy can be approached by combining Model Predictive Control(MPC) with accurate forecasts (e.g. of the power demand). Wenow review some approaches that have been proposed in theliterature. Gear shifting.
Gear shifting control is available in automatedtransmissions, and impacts the way the upstream powertraincomponents are operated: in vehicles with manual transmis-sion, it is commonly advised to “up-shift soon”, which trans-lates into operating the engine at low speed and high torque,where e ffi ciency is usually higher. We also know that this ispossible only to some extent, because drivability (i.e. the re-sponsiveness of the vehicle to our inputs) is adversely a ff ected.Production gear shifting controllers are generally rule-based;extensive testing and tuning can deliver good fuel economy anddrivability [39].If the future wheel speed and torque can be predicted reliably,gear shifting control can be formalized as an optimal controlproblem and solved by various techniques. Since most trans-missions only feature a finite number of gears, the system dy-namics are discrete. In [62, 63, 64], the gear shifting problem issolved jointly with the energy management problem, combin-ing dynamic programming with Pontryagin’s minimum princi-ple in [62], and with convex optimization in [63]. In [64], alsoengine on / o ff is included; the resulting mixed integer non-linearprogram is treated as a distributed optimization problem, andreformulated as a two-layer MPC problem. [65] uses the min-imum principle and dynamic programming to jointly solve thegear shifting problem and the longitudinal control problem, foran fuel-powered vehicle.A simplified problem can be obtained assuming that thetransmission gear ratio domain is continuous; in practice, this6s only true when the vehicle is equipped with a continuouslyvariable transmission. Even when this is not true, one can geta suboptimal solution by rounding the optimal gear ratio to thenearest available value [66]. Engine on / o ff . Engine control includes a vast family of chal-lenging problems, such as knock control, air / fuel ratio control,thermal control (see e.g. [67] on the topic). Engine on / o ff con-trol determines whether to idle or shut fuel injection o ff . Atrivial approach to the problem is to cut injection as soon as thepower request is non-positive (in human-driven vehicles, whenthe gas pedal is released); this causes a sudden reduction oftorque and, ultimately, vibrations and discomfort. From a fueleconomy perspective, restarting the engine has a cost, that isgenerally lower than the cost of a cold start, but may be higherthan the cost of idling for a short time. Still, in favorable con-ditions and with a su ffi ciently long preview of the upcomingdriving profile, fuel savings between 5 and 10 % were reported[68, 69].The engine on / o ff control problem is studied in [70] for aconventional powertrain, using a hybrid systems formulation;control design considers a relaxation to the continuous domain,and maps the solution back into the hybrid domain. The sameproblem is studied in [71, 72] for a belted starter alternator in ahybrid electric vehicle, with the main focus being on vibrationand noise reduction. A similar setup has been considered inseveral other works, where the engine on / o ff and energy man-agement problems are solved jointly [73, 63, 74]; in this case,the engine mode is often determined by dynamic programming. Energy management.
By energy management we refer to theproblem of allocating, in hybrid vehicles, the power demandto the internal combustion engine and the electric motor. Thisproblem has been extensively studied in the literature: we re-fer the interested reader to [75, 76, 77] for extensive literaturereviews and to [78, 79] for systematic comparisons between ex-isting approaches.In an optimal control formulation, the limited energy stor-age capability of batteries can be translated into a terminal stateconstraint, see e.g. [80, 81]. In hybrid electric vehicles, thebattery cannot be recharged from the grid, therefore the termi-nal battery charge is often constrained to its initial or nomi-nal value. If the driving schedule is known in advance, thisproblem is easily solved by dynamic programming [82, 83].The so-called Equivalent Consumption Minimization Strategy(ECMS) can be derived from Pontryagin’s minimum principleand the observation that (under certain modeling assumptions)the adjoint state λ (roughly speaking, the Lagrange multiplierassociated to the terminal battery charge constraint) is constantfor a fixed driving cycle; the optimal trajectory is found by it-eratively determining the optimal λ . The reader is referred to[84, 85, 86, 87] for details on the model assumptions, guaran-tees of optimality, implementation details, and performance incase the assumptions are violated.In plug-in hybrid electric vehicles, the battery charge fullyutilized, hence the trade o ff between fuel and electricity con-sumption leads to an interesting optimization problem [88]. ECMS approaches for plug-in hybrids are summarized in [89].A key aspect is the discharge rate of the battery; ideally, the bat-tery is gradually discharged and reaches the minimum chargeonly at the end of the trip. This requires route information andlong-term planning, and is discussed in Section 5.1Some real-time approaches borrow the ECMS formulation;if the driving schedule is not known in advance, various up-date laws for λ have been proposed, based on historical dataand forecasts [90, 91]. The generation of the reference stateof charge, discussed in Section 5.1, plays an important rolein this regard. Approaches that systematically address the in-formation gap in real-time are mostly based on robust control[78, 92], stochastic dynamic programming, and MPC. Stochas-tic dynamic programming is used e.g. in [93] to minimize thediscounted infinite-horizon cost, and in [94, 95] in a shortestpath formulation. A stochastic optimal control framework isdeveloped in [96, 97] to determine the policy minimizing thelong-run expected average cost. All formulations yield a causal,time-invariant, state-feedback controller that can be fairly eas-ily implemented.MPC provides a systematic framework to include forecastsand handle constraints in real-time. The authors of [98, 99]discuss a nonlinear MPC approach, in which an approximationof the cost-to-go is derived using the relationship between dy-namic programming and Pontryagin’s minimum principle. Asimilar approach is proposed in [100], where a preview of fu-ture velocity is exploited. [101] is focused on the velocity pre-diction for MPC-based energy management. [102] proposes astochastic MPC approach, modeling the power demand fromthe driver as a Markov chain and training it using standard driv-ing cycles and historical driving data. [103, 104] extend thisapproach showing how the driver model can be learned online.Although often disregarded in the scientific literature, auxil-iary devices like air conditioning and lights can have a majore ff ect on energy consumption; to some extent, they can also becontrolled. For example, the air conditioning may be adjustedto preserve the electric driving range [105].Instead of minimizing only energy consumption, several au-thors have addressed also di ff erent optimization goals, such aspointwise powertrain e ffi ciency [106], drivability [107], pollu-tant emissions [63, 108], battery aging [109, 110, 111], drivingcost [112, 113, 114, 115]. The MPC approach in [116], in-stead, combines longitudinal control and energy management,exploiting forecasts of tra ffi c signals and road slope. Challenges and opportunities for CAVs
Gear shifting control, engine on / o ff control, and energy man-agement are generally aimed at minimizing the cost function inequation (1), and can benefit from forecasts of the vehicle speedand of the torque or power demand. These algorithms can nat-urally be integrated, to better manage the powertrain and theassociated uncertainty [64]. In CAVs this opportunity can becombined with more reliable forecasts. In facts, the future pro-files of vehicle velocity, wheel torque and power demand canbe (to some extent) predicted, because of • driving automation and the removal (or substantial reduc-7ion) of unpredictable human factors; • the awareness of the surrounding environment due to per-ception sensors and communication with other vehiclesand infrastructure.In gear-shifting and engine on / o ff control, this opportunitymostly relates to the avoidance of energy-wasteful events: ev-ery switching and shifting has a cost, and switching decisionsare intrinsically reliant on forecasts or assumptions on the fu-ture. In energy management, we have documented how recentresearch has focused on filling the information gap on the futuredemand. Both short-term forecasts (as the ones just discussed)and long-term forecasts (which are handled as described in Sec-tion 5.1) carry valuable information in this sense. Example: MPC approach for a plug-in hybrid electric vehicle
In a plug-in electric vehicle, powertrain control includes gearshifting, engine on / o ff , and energy management. We formulateit as the following finite-horizon optimal control problem in thetime domain.minimize u | t , u | t ,..., u N − | t N − (cid:88) k = g ( x k | t , u k | t , w k | t ) + l ( x N | t )subject to x k + | t = f ( x k | t , u k | t , w k | t ) , = h ( x k | t , u k | t , w k | t ) , u k | t ∈ U ( w k | t ) , x k | t ∈ X , k = , . . . , N − , x | t = x t , x N | t ∈ X N . (2)Let (cid:104) u ∗ | t , u ∗ | t , . . . , u ∗ N − | t (cid:105) be the solution at time t = t . The firstinput u ∗ | t is applied, and at the next time step t = t + T s theoptimal control problem is solved using the new measurements x t . The MPC control law is u t = u ∗ | t .We set the state vector to x = [ E q , n g , s e ] T , the input vectorto u = [ T m , T e , u g , u e ] T , and the forecast vector to w = [ v , P a ] T ,where E q is the energy stored in the battery, n g is the gear num-ber, s e is the engine on-o ff state, T m is the motor torque, T e isthe engine torque, u g is the gear shifting command, u e is theengine on / o ff command, v is the vehicle longitudinal speed, P a is the power consumption of electric auxiliaries.We model the powertrain dynamics as in [74] and we applyEuler discretization with step T s , obtaining f ( x , u , v ) = E q − T s A b R b Q b (cid:18) E q − (cid:113) E q − R b Q b A b P b E q (cid:19) n g + u g s e + u e , (3)where A b and B b fit the battery open circuit voltage, R b is thebattery internal resistance, Q b is the battery capacity. The alge-braic constraint h enforces the summation of T m and T e at thetransmission input shaft, and the summation of motor power P m and auxiliary power P a at the battery output, h ( x , u , w ) = (cid:34) T t ( v , n g ) − T m − s e T e P b − P m ( v , T m ) − P a (cid:35) . (4)The input transmission torque T t is determined from a vehiclelongitudinal model and from the transmission gear ratio; here we have implicitly assumed that T t is a known nonlinear func-tion of v and n g . The same can be said for the motor speed;therefore, the motor power P m is a known nonlinear function of v and T m . We wish to minimize the total powertrain energy g ( x , u , w ) = γ f P f ( n g , s e , T e , v ) + γ q P q ( n g , T m , v ) , where P q = f ( x , u ) − E q and P f is a nonlinear mapping from theengine speed and torque to the fuel thermal power; the mappingfrom v to the engine speed is implicitly embedded.The input constraint set U ( v ) = U ( v ) × U ( v ) × U × U defines the actuator limits, where U ( v ) = (cid:110) T m : T m ( v ) ≤ T m ≤ T m ( v ) (cid:111) , U ( v ) = (cid:110) T e : T e ( v ) ≤ T e ≤ T e ( v ) (cid:111) , U = (cid:110) u g : u g ∈ {− , , + } (cid:111) , U = { u e : u e ∈ {− , , + }} . T m , T m , T e , T e are nonlinear functions of the motor and enginevelocities, and their mapping to the vehicle speed v is implicitlyembedded. The state is confined to a safe operating region forthe state of charge (to avoid overcharge or overdischarge), andto the discrete domains of the gear number and engine state, X = X × X × X , where X = (cid:110) E q : E q ≤ E q ≤ E q (cid:111) , X = (cid:110) n g : n g ∈ (cid:110) , , . . . N g (cid:111)(cid:111) , X = { s e : s e ∈ { , }} . The terminal battery charge must exceed a reference value, X N = (cid:110) E q : E (cid:63) q ≤ E q (cid:111) ; in our architecture, E (cid:63) is a position-dependent reference that is computed remotely in the chargeplanning block (described in Section 5.1). In closed loop, thisconstraint a ff ects the actual battery charge at the end of the trip,which in turn a ff ects the recharge time, i.e. the minimum waituntil the next trip. The terminal cost l is another knob thatcan improve closed loop performance in the long term; if itapproximates the optimal cost-to-go su ffi ciently well, it helpsthe MPC policy to approach the optimal infinite horizon pol-icy [117]. In this application, a ffi ne approximations of the form l = a + ( E q − E (cid:63) q ) b have been shown to give good results [99].We refer to [62, 64] for numerical techniques to solve prob-lem (2) and simulation analysis of the closed loop performance. Motion control ensures that the vehicle’s longitudinal and lat-eral motion follows a reference trajectory or path. A simplelongitudinal control is cruise control, which tracks a constantreference velocity specified by the driver. Next we review themain control systems for longitudinal and lateral motion.
Literature review
We first organize the existing longitudinal control approachesby their use of external information: predictive cruise control(using a reference velocity computed remotely), adaptive cruise8ontrol (adjusting the reference velocity based on the percep-tion data), urban cruise control (using communication with theinfrastructure), cooperative adaptive cruise control (using com-munication between vehicles). We then move to lateral control.
Predictive cruise control.
By predictive cruise control, we in-dicate a cruise control tracking a reference velocity that is gen-erated using preview information [118, 119]; information canbe static (like road grade and speed limits) or dynamic , butslowly changing (like tra ffi c speed). As such, the reference tra-jectory generation is often cloud-aided (i.e. it exploits informa-tion that is generally retrieved from the cloud) and can be castas an optimization problem. A closely related problem is eco-driving, which is concerned with velocity trajectory optimiza-tion for minimum energy consumption; this aspect is discussedin Section 5.2. This reference trajectory is based on long-termforecasts and cannot be implemented in open loop. The real-time control simply tracks the reference signal, does not exploitperception sensors or cooperation, and requires driver interven-tion to ensure basic safety. Nonetheless, reference generationfor predictive cruise control can also be integrated with any ofthe advanced cruise controls discussed next. Adaptive cruise control (ACC).
ACC is an enhanced cruisecontrol, which detects any preceding vehicle and adjusts speedin order to avoid collisions [120, 121]. ACC design is orientedto the enhancement of passenger safety and comfort, and tobroader impacts like improved road throughput and energy ef-ficiency.MPC has proven e ff ective in simultaneously guaranteeingACC safety and performance. In MPC, safety in closed loopis closely related to the problem of persistent feasibility [122],which is related to the choice of the terminal cost and con-straints; choosing the terminal set as a control invariant set canensure stability and persistent feasibility. Computing a controlinvariant set is not trivial in the presence of nonlinear dynam-ics and time-varying, non-convex state constraints. In ACC,a conservative approximation is to assume that the precedingvehicle can fully brake at any time; in practice, this can turnout to be too conservative. Also notice that the preceding ve-hicle forecast is uncertain; while certainty equivalence can beadopted, robust and stochastic formulations may be more sys-tematic. We refer to [123, 61] for a more detailed discussionon this. An important role is also played by the inter-vehicularspacing policy: most systems adopt a constant distance policyor a constant heading time policy [121, 124, 125], as we dis-cuss further in the next paragraph. While guaranteeing safety,various performance objectives can be pursued, such as roadthroughput [121], fuel economy [124, 126], and driver comfortby mimicking her driving style [123].In applications oriented to energy e ffi ciency improvement, acommon approach is to pursue a small inter-vehicle gap; at highspeed, this can reduce the aerodynamic resistance. In open-roadexperiments with a platoon of trucks, fuel reductions up to 7 %have been registered [127]; for the case of compact vehicles,a study with one-eighth-scale models showed considerable re-duction of fuel consumption [128]. Combinations of ACC and predictive cruise control are also often proposed. More pre-cisely, a long-term reference velocity is computed based onstatic and slowly changing information (as discussed in Sec-tion 5.2); safety in closed loop is guaranteed tracking this refer-ence with an ACC [118]. Urban cruise control.
V2I communication can provide lookahead information about tra ffi c and signalized intersections inthe downstream road. The strategies to explore this informa-tion are well addressed, especially in arterial scenarios wherethe vehicle is driving in tra ffi c through a series of tra ffi c lights.When the vehicle receives signal phase and timing information,MPC strategies for ACC have shown substantial energy saving[118, 129, 116]. Compared to a standard ACC, the signal in-formation introduces additional position-dependent constraints,to enforce that the downstream intersection is crossed during agreen phase. Generally, the MPC implemented on-board has alimited prediction horizon, both because the V2I communica-tion range is limited, and to reduce the computational burden.As discussed in Section 5.2 it is possible to use statistical andhistorical signal data to (remotely) compute a reference velocitywith a long horizon. This reference velocity can be tracked bythe on-board urban cruise control, which ensures safety usingthe real-time perception and V2I data. Cooperative Adaptive Cruise Control (CACC).
CACC is anenhancement of ACC enabled by communication. The per-formance of ACC is limited by perception systems, that (evenin the absence of noise and delays) can only measure the rel-ative distance and velocity. V2V communication enables theexchange of vehicle acceleration (and potentially of its fore-cast), which can be extremely valuable in dynamic driving sce-narios. CACC can exploit this additional piece of informa-tion to guarantee higher safety and smaller inter-vehicular dis-tances [130, 131, 132, 133]. In addition to improved safety, thiscan translate into lower energy consumption [134], higher roadthroughput [130], and passenger comfort: in [135], passengersusing a CACC were found to be comfortable with inter-vehicletime gaps between 1 s and 0 . • Dynamics , i.e. the dynamics of each CAV. • Information flow network , i.e. the topology and quality ofinformation flow, and the type of information exchanged.Figure 4 depicts some typical communication topologies9 a) Predecessor Following. (b) Bidirectional.(c) Predecessor Following Leader. (d) Bidirectional Leader.(e) Two Predecessor Following. (f) Two Predecessor Following Leader.Figure 4: Information flow topologies in a four vehicle platoon. The red nodeindicates the platoon leader. The nomenclature is taken from [136]. used in platooning. The information exchanged may justbe the current velocity and acceleration, or include fore-casts thereof and information on lateral motion. • Local controller design and its use of on-board informa-tion. • Formation geometry , i.e. vehicle ordering, cruising speed,and inter-vehicle distance.Notice that the dynamics and the distributed controller pertainto the individual CAV, while the information flow network andthe formation geometry are properties of the platoon. The lat-ter two can be decided a priori in a specific demonstration, butrequire some form of standardization for operation on publicroads. A possibility is to coordinate remotely the informationflow network (based on the instrumentation and on the num-ber of vehicle involved) and the formation geometry (based onvehicle characteristics, origins and destinations). We furtherdiscuss this point in Section 5.3.To ensure safety in closed loop, each CAV must be stablewith su ffi cient robustness margins. In an MPC setting, safetycan be addressed as in ACC [139], although with reduced con-servatism thanks to V2V communication. However, distur-bances acting on the platoon leader may still be amplified inthe downstream vehicles; this phenomenon is known as stringinstability [140, 141]. String stability is a property of the localdistributed controller, but it has been shown to depend on theinformation flow network and the formation geometry [141]. Acrucial aspect is the inter-vehicle spacing policy, i.e. the choiceof a reference relative distance d (cid:63) between two consecutiveCAVs [142, 143]. Most works in the literature adopt a simpleconstant distance policy d (cid:63) = d (cid:63) or a constant time headwaypolicy d (cid:63) = t (cid:63) h v + d (cid:63) , where t (cid:63) h is the constant time headway from the preceding ve-hicle, v is the current speed ego CAV, and d (cid:63) is a constant min-imum distance.Looking at the overall system performance and broader im-pact of platooning, the inter-vehicular spacing or heading timeis generally regarded as the main metric. A minimum gap max-imizes the road throughput [134, 133, 144, 145] and can re-duce the vehicle air drag [130, 146]. Recent studies on CACC for energy e ffi ciency have highlighted an inherent trade o ff be-tween air drag reduction (via reduced inter vehicular distance)and powertrain e ffi ciency. More precisely, this trade o ff is likelyto be significant when the velocity profile is variable: maintain-ing a small gap may require aggressive throttling and braking,and may lead to suboptimal operation of the powertrain. In[144], this problem is studied for heavy duty vehicles, whenthe speed variability is due to road grade; the proposed solu-tion includes a centralized high-level (cloud-based) generationof a speed reference, and a decentralized vehicle-level track-ing controller; similarly to [123], robust invariance is used toensure closed loop safety in the CACC. In [126, 145], the prob-lem is approached for light duty CAVs, using forecasts of thepreceding vehicle’s velocity; di ff erent MPC formulations arepossible, depending on the availability of a powertrain model.Other CACC approaches for energy e ffi ciency were presentedin [147, 148, 149, 150].CACC is fundamentally a tracking control problem withforecast; as such it has been addressed with a variety of con-trol techniques [136]. Linear consensus control and distributedrobust control techniques enable insightful theoretical analysisand can provide guarantees of string stability [151, 152]; a pit-fall is the limitation of the dynamics to the linear domain, andthe lack of guarantees in the presence of constraints. MPC canincorporate nonlinear dynamics, input and state constraints, andforecasts [153, 144, 154, 145]. A distributed MPC formulation,suited for any information flow topology, has been presented in[155]. Lateral vehicle control.
Lateral control supervises the vehiclemotion in the lateral direction, actuating the steering angle ortorque. Generally, lateral controllers track a reference trajec-tory or path from the motion planning block (described in Sec-tion 4.3), ensuring safety and robustness to model uncertaintyand a fast changing environment.MPC has been fruitfully employed for lateral control, dueto its ability to handle constraints and complex vehicle dynam-ics; for example, a nonlinear bicycle model was used in [156].[157] presents an MPC for integrated longitudinal and lateralcontrol using a linear time-varying model. The MPC-basedlateral control in [158] uses a linearized conservative lateraldynamics model and a overreacting lateral dynamics model toaccount for two extreme cases in lateral cornering. In [159],a piece-wise a ffi ne model is used for trajectory stabilization inthe active steering system.Some works are specifically focused on the lateral controlof CAVs. The lateral controllers in [160, 161] track the lateralmotion of the platoon leader. In [162], an MPC-based lateralcontroller uses vehicular communication to enhance safety inmotion planning and control.Lateral control design is deeply intertwined with motionplanning; in fact, both algorithms are often based on the samemodels and measurements. Forecasts from communication af-fect lateral motion also through the motion planning block, asdiscussed in Section 4.3.10 hallenges and opportunities for CAVs In the deployment of cooperative driving controls, knownchallenges include the diversity of communication topologiesand protocols, communication delays, packet losses, and com-plex dynamics. While progress has been made to systematicallyanalyze these complex and heterogeneous systems (at least inthe highway platooning case), a comprehensive framework isstill lacking at present.Similarly, the trade o ff between safety and robustness re-quirements (like string stability) and broader impacts (like en-ergy consumption and road throughput) has been partially stud-ied, but a comprehensive analysis has not emerged yet; for in-stance, the value of forecasts in this trade o ff is not yet entirelyclear. In CACC, most often the preceding vehicle communi-cates its current velocity and acceleration; however, V2V com-munication allows extended forecasts that, although not per-fect, may be helpful in reducing conservatism. The balance be-tween communication bandwidth and closed loop performancehas not been thoroughly addressed. Example: MPC for cooperative adaptive cruise control
A CAV can implement longitudinal control in virtually anydriving scenario, including highways, urban roads and ruralroads. We formulate the problem in the time domain, as a finite-horizon optimal control problem of the following form.minimize u | t , u | t ,..., u N − | t N − (cid:88) k = g ( x k | t , u k | t , w k | t ) + l ( x N | t )subject to x k + | t = f ( x k | t , u k | t , w k | t ) , u k | t ∈ U , x k | t ∈ X , k = , . . . , N − , x | t = x t , x N | t ∈ X N . (5)Let (cid:104) u ∗ | t , u ∗ | t , . . . , u ∗ N − | t (cid:105) be the solution at time t = t . The firstinput u ∗ | t is applied, and at the next time step t = t + T s the opti-mal control problem (5) is solved using the new measurements x t . The MPC control law is u t = u ∗ | t .We set the state vector as x = [ d , v ] T , the input vector as u = [ F w , F b ] T and the forecast vector as w = v (cid:63) p , where d is thedistance to the preceding vehicle, v is the vehicle speed, F w isthe wheel force, F b is the braking torque, v (cid:63) p is the velocity ofthe preceding vehicle. We model the longitudinal dynamics asin [39] and apply Euler discretization with step T s , obtaining f ( x , u ) = (cid:34) d + T s (cid:0) v (cid:63) − v − L (cid:1) v + T s M (cid:16) F w − F b − F f (cid:17)(cid:35) , where M is the vehicle mass, L is the vehicle length, and F f = Mg sin ϑ − Mg ( C r + C v v ) − ρ AC x v , g is the gravity constant, ϑ is the (position-dependent) roadslope, C r is the rolling coe ffi cient, C v is the viscous frictioncoe ffi cient, ρ is the air density A is the front area, C x is the airdrag coe ffi cient. The stage cost g ( x , u ) is a trade o ff between thecontrol e ff ort, the velocity tracking error v (cid:63) − v , and the distancetracking error d (cid:63) − d . If the CAV has free road ahead, the refer-ence velocity v (cid:63) and distance d (cid:63) are defined by the eco-driving block described in Section 5.2; otherwise, v (cid:63) is dictated by thepreceding vehicle, v (cid:63) = v (cid:63) p .The input constraint set U defines the actuator limits. Weconstrain speed and acceleration to a convex set X , and en-force collision avoidance in the prediction horizon by a constantminimum distance, X = (cid:110) d : d ≥ d (cid:111) . Distance d is measuredonly at time t , and evolves in the prediction horizon accord-ing to the system dynamics and to the velocity of the precedingvehicle v (cid:63) p . In adaptive cruise control, only v (cid:63) p , t is known andthe prediction along the horizon must be based on some model[163, 139, 123, 61]. In cooperative adaptive cruise control,also ˙ v (cid:63) p , t is known; potentially, the preceding vehicle can sharea forecast of its future acceleration, (cid:20) ˙ v (cid:63) p , | t , ˙ v (cid:63) p , | t , . . . , ˙ v (cid:63) p , N f − | t (cid:21) ,where N f is the forecast horizon. In an emergency braking sce-nario, the CAV must come to a complete stop to avoid colli-sion with a static obstacle (like a stopped vehicle) with forecast v (cid:63) k | t = , ∀ k = , , . . . , N −
1. This approach can also be used toenforce safe crossing of intersections: stop signs and red lightscan simply be treated as static obstacles. In sum, the state con-straint set is given by X = X ∩ X , and is dynamically shapedby the current measurements and forecasts.In [126], we used the formulation (5) and used a forecastwith N f >> N to compute a terminal set X N with the followingproperty: if x N | t ∈ X N , then the ego-vehicle can avoid colli-sions with the preceding vehicle ( x k | t ∈ X ) without applyingany braking force ( F b , k − | t =
0) throughout the forecast hori-zon ( k ∈ [ N + , N f ]). Figure 5 shows an experimental resultobtained with this approach. The preceding vehicle followsa sinusoidal velocity profile; the ego vehicle maintains a safedistance without applying any hard braking. It can be notedthat the relative distance does not reach the allowed minimumvalue, but rather oscillates around the desired distance d (cid:63) toavoid power dissipation through braking. We refer to [126] forfurther analysis and details about the implementation. The real-time motion planning block generates a referencetrajectory for the longitudinal and lateral motion of the CAV. Itfollows higher level specifications, namely the waypoints de-fined by the eco-routing block (as described in Section 5.4) andthe recommended speed by the eco-driving block (describedlater in Section 5.2) or the commands from the multi-vehiclecoordination block (which can include speed recommendationand distance from other vehicles, described later in Section 5.3).While these high-level references are computed as a functionof the overall trip, the real-time motion planning block has in-formation on the actual state of the surrounding environment,based on the perception sensors and on communication. It isresponsible for finding a trajectory that respects driving rulesand is feasible for the lower level controllers, comfortable forthe passengers, and in line with the high level directions.
Literature review
The literature on motion planning is vast and covers a widespectrum of applications and computational techniques. Com-prehensive reviews can be found in [164, 59]; here we only give11
100 200 300 400050100150200 d i s t an c e ( m ) desired distanceminimum distanceactual distance time (s) v e l o c i t y ( m / s ) preceding vehicleego vehicle
150 200102030150 200152025
Figure 5: Catch-up of a vehicle traveling with sinusoidal velocity profile: inter-vehicle distance and velocity. Zoomed portions shows the behavior during oneperiod of the sinusoidal profile. a brief overview of the problems and the related computationaltechniques. The decision making process includes both deci-sions on the vehicle behavior (stay in the current lane or change,come to a stop at an intersection, yield to pedestrians, etc.) andon how to translate that behavior into a CAV trajectory.Decision making for vehicle behavior can be implementedby heuristic rules, although on public roads the intentions ofother agents (aside from other CAVs) are uncertain. As a con-sequence, estimation, prediction, and learning techniques playa major role in this field [59].The high level specifications from the remote planning andthe behavioral decision must then be translated into a pathor trajectory for the CAV, taking into account the most re-cent state and prediction of the surrounding objects. In theliterature, path planning generally indicates the problem ofplanning future motion in the configuration space of the vehi-cle, while trajectory planning indicates the search of a time-parametrized solution [59, 164]. Both problems are often for-mulated as an optimization problem, i.e. in terms of the mini-mization of some cost functional subject to constraints. Holo-nomic constraints include collision avoidance constraints andterminal constraints, that define safe regions or driving corri-dors from the current to the target configuration, in the exis-tence of static obstacles, road boundaries, tra ffi c rules. In pathplanning, di ff erential constraints are included to enforce somelevel of smoothness in the solution, for instance on the pathcurvature. In trajectory planning, dynamic constraints can beincluded, and collision avoidance can in principle be enforcedalso for dynamic (not only static) obstacles.Optimal path and trajectory planning are both PSPACE-hard in their general formulations [59]. As such, past and currentresearch have focused on computationally tractable approxima-tions, or on methods that apply to specific scenarios. Compu-tational methods for path planning include variational methods,graph search methods, and incremental search methods. Tra-jectory planning can be solved using variational methods in thetime domain, or converting it to a path planning problem with atime dimension [59].In CAVs, communication can greatly improve the awarenessof other agents and, in general, of the surrounding environment.In fact, CAVs can share not only their own states and predictedmotion, but also the obstacles that they detect; a distributed per-ception system can include multiple CAVs, road-side units, andpotentially cyclists and pedestrians, and can significantly out-perform an advanced perception system that only uses on-boardsensors. This potential is shortly discussed in [164]. Some ex-amples of motion planning applications using forecasts fromcommunication can be found in [165, 162, 161]. A strictly re-lated branch of literature is on the coordination of CAVs, forinstance for the case of an autonomous intersection. We willdiscuss further these applications in Section 5.3. Challenges and opportunities for CAVs
As anticipated, the variety of subproblems and of compu-tational techniques is large. In general, a motion planning ap-proach needs to be determined based on the specific application,including the type of vehicle and the environmental constraintsthat it is likely to encounter. There exists a trade o ff betweenthe complexity of the motion planning block and that of thelongitudinal and lateral control blocks. If the motion planningblock accounts for an accurate dynamic model of the CAV, thenthe motion controls may be simplified.The other approach, i.e.using simplified motion planning and sophisticated motion con-trols, is also possible. At present, a systematic framework forsystem designers to allocate complexity to the di ff erent func-tional blocks has not emerged.Strictly related to this is the interaction with the perceptionsystem. The current map and prediction of the surrounding en-vironment has a crucial e ff ect on decision making and motionplanning; in some approaches, models of other agents are inte-grated in the decision making process. Opportunities for CAVslie in their extended sensing capability: there is a variety ofscenarios in which V2V and V2I communication give a crucialadvantage over the most advanced perception systems. Suchscenarios include obstacles outside the line of sight, and drivingat small distance from the preceding vehicle. Motion planningtechniques for CAVs driving on public roads that exploit thisadvantage still need to be completely exploited.With CAVs, we are not only interested in the safe motion of asingle agent, but potentially to the safe and coordinated motionof multiple agents; an example that has been touched previouslyis that of platooning. In multi-agent scenarios, motion planningneeds to address new problems (like platoon merging and dis-mantling), and to solve existing problems (like lane changing,and collision avoidance) in a way that is feasible for the wholeformation.12 xample: collision avoidance We now show an optimization formulation for collisionavoidance that has been recently proposed in [166]. For sim-plicity, we focus on the case of a single CAV moving in a 2-dimensional space while avoiding multiple obstacles. We for-mulate an optimal trajectory planning problem, which can di-rectly consider account for the system dynamics and actuatorlimits. We formulate the problem as follows.minimize u ,..., u N − N − (cid:88) k = g ( x k , u k )subject to x k + = f ( x k , u k ) , u k ∈ U , x k ∈ X , E ( x k ) ∩ O m = ∅ , k = , . . . , N − , m = , . . . , M , x = x t , x N ∈ X N . We set the state vector as x = [ X , Y , ψ, v ] T and the input vectoras u = [ a , δ ] T , where X and Y are the coordinates of the vehiclecenter of mass in an inertial frame, ψ is the inertial heading, v is the longitudinal velocity, a is the longitudinal acceleration, δ is the steering angle. The system dynamics f are given by thekinematic bicycle model [167, 159]˙ X ( t ) = v ( t ) cos ( ψ ( t ) + β ( t )) , ˙ Y ( t ) = v ( t ) sin ( ψ ( t ) + β ( t )) , ˙ ψ ( t ) = v ( t ) l r sin ( β ( t )) , ˙ v ( t ) = a ,β ( t ) = tan − (cid:32) l r l r + l f tan( δ ( t )) (cid:33) , after forward Euler discretization. The convex sets U and X model the actuator and speed limits. The cost function l is takenas a weighted sum of the time and input e ff ort.The main challenge is represented by the collision avoidanceconstraints E ( x k ) ∩ O m = ∅ , where E ( x k ) denotes the spaceoccupied by the CAV and O m are the obstacles to avoid; in gen-eral, these constraints are non-convex and non-di ff erentiable.In [166], such constraints have been reformulated into smoothnonlinear constraints. The reformulation is non-conservativeand can be applied to problems where the CAV and the obsta-cles can be represented as a finite union of convex sets. Werefer the reader to [166] for results on an autonomous parking;the problem formulation is analogous to the one presented here,except minor di ff erences in the system dynamics (accountingfor the fact that parking maneuvers happen at low speed).
5. Remote planning and routing
In Figure 2, the remote planning and routing blocks performlong-term computations to exploit route and tra ffi c data, andmaximize the CAV overall trip performance. Metrics includevehicle energy consumption, trip time, driver convenience, androad throughput. Coordination with the on-board functionalblocks is fundamental to obtain the desired performance im-provement. In this section, we review the existing literature for each ofthe three functional blocks in the on-board layer of Figure 2: battery charge planning , eco-driving , eco-routing . The sepa-ration in blocks helps organizing our review, but practical ap-proaches often trespass these boundaries. For the sake of clar-ity, we discuss eco-driving approaches for isolated vehicles andfor groups of vehicles in two separate paragraphs. Our goal isto survey the existing literature, determine the potential of real-time access to data and remote computations, and to presentsome selected approaches.Many of the algorithms discussed in this section here canbe implemented in the cloud. In some cases, the algorithmsmay be implemented on a CAV (the ego or another one) or in aroad-side coordinator; we will highlight these cases in the dis-cussion. While reference generation and routing may be per-formed only once (at departure), in most cases re-computationalong the trip is advised or required; re-computation may beperiodic or event-based periodic, depending on the application. Despite the advances in battery technology, driving rangeand charging time are still pressing problems in any electrifiedpowertrain. Route data and tra ffi c and weather forecasts candramatically improve the accuracy of the electric driving rangeestimate. More specifically: • In electric vehicles, battery depletion can be more accu-rately predicted, and potentially counteracted by limitingthe auxiliaries power, the traction power to the driver [168]or planning stops at charging stations. • In Hybrid Electric Vehicles (HEVs), the charge should re-main bounded throughout the trip, and equal a target valueat the end. Most real-time energy management approaches(including those described in Section 4.1) postulate a ref-erence charge signal, which is usually chosen constant.This simple choice is also logical in the absence of infor-mation, but it may make it di ffi cult to satisfy the chargeconstraints, for instance if the trip includes large altitudevariations [91]. • In plug-in HEVs, the final charge should be greater orequal to a minimum level (chosen to ensure that the bat-tery does not incur deep discharge). A simple strategythat is widely used in practice is the so-called Charge-Depleting / Charge-Sustaining (CDCS) strategy [169, 170].The battery is (on average) discharged during the chargedepleting phase; in the extreme case, the plug-in HEV isoperated as an electric vehicle and the discharge rate ismaximum. During the charge sustaining phase, the chargeis kept (on average) constant and the plug-in HEV is op-erated as an HEV. The strategy is conceptually simple butusually suboptimal.In [91], the reference charge trajectory for an HEV is com-puted using the elevation profile and the speed limits (or aver-age tra ffi c speed). The main goal is to keep the battery charge13ithin the prescribed limits throughout the trip, and to maxi-mize the energy recuperation during deceleration and downhillsegments. The two goals are intertwined, especially in HEVswith small batteries: the battery charge needs to be dynamicallycontrolled to full exploit recuperation. [171] uses a similar ap-proach for a plug-in hybrid electric bicycle.In [172, 173], two approaches to compute the reference stateof charge of a plug-in HEV. Both approaches use logged dataof velocity and altitude on a given route (that is assumed tobe a commuting route). The first approach computes a refer-ence state of charge trajectory by solving a convex program;the second approach determines an optimal cost-to-go functionby dynamic programming. The two approaches yield very sim-ilar performance and clearly outperform the CDCS approach.A similar problem is considered in [114, 115], for a modular plug-in hybrid electric vehicle, in which an engine and a gen-erator are mounted on a trailer that can be detached from themain electric vehicle. In the analyzed scenarios, the trailer canbe rented at fixed locations along the route, with di ff erent fueland pricing options. The optimal solution includes both the op-timal trailer rental policy and the optimal battery dischargingpolicy. Battery charge trajectory planning for plug-in HEVs isalso studied in [101] based on real-time tra ffi c data. A compu-tationally e ffi cient, yet meaningful model of the plug-in HEV isspecifically developed for this purpose. The planner generatesa battery charge trajectory that is used in real-time as a terminalstate constraint. [174] also pursues the battery charge trajec-tory planning for a plug-in HEV; unlike the works above, thatexploit various levels of route information, this work only as-sumes that a sampled probability distribution of the trip length(extracted from past trip data) is known a priori. Challenges and opportunities for CAVs
The problem of battery charge planning strongly relies onprior information on the future driving schedule. An integratedCAV control architecture enables the access to extended and ac-curate forecasts. While most existing approaches exploit staticroute information (like road grade and speed limits), in an in-tegrated CAV architecture one can accurately predict the futurepower demand from the eco-driving block (discussed in Sec-tions 5.2 and 5.3), and accordingly optimize charge depletion.The same framework could include forecast uncertainty, or op-timize jointly the battery charge and the vehicle velocity trajec-tories.Other opportunities for electric and plug-in hybrid vehicleslie in the interaction with the electric grid. Planning of stops atcharging stations can be included in the planning problem, in-cluding charging and waiting times, dynamic pricing, and non-trivial models of charging. Other related topics are grid balanc-ing and the interactions with smart grids [47, 48, 49].
Example: battery charge planning for a connected plug-in hy-brid electric vehicle
We formulate the battery charge planning problem in the timedomain, as a finite-horizon optimal control problem of the fol- lowing form.minimize u , u ,..., u N − N − (cid:88) k = g ( x k , u k , w k )subject to x k + = f ( x k , u k , w k ) , = h ( x k , u k , w k ) , u k ∈ U ( w k ) , x k ∈ X , k = , . . . , N − , x = x t , x N ∈ X N . (6)We set the state vector to x = E q , the input vector to u = [ T m , T e ] T , and the forecast vector to w = [ v , P a ] T where E q is the battery internal energy, T m is the motor torque, T e isthe engine torque, v is the vehicle longitudinal speed, P a is thepower consumption of electric auxiliaries. The forecast of v can simply be the reference speed generated by the eco-drivingblock (see Section 5.2 and Section 5.3). The forecast of P a may be produced using weather forecasts and a model of theon-board air conditioning, assuming the latter is the main causeof power consumption.We model the powertrain dynamics as in (3) but, for sim-plicity, we do not optimize the gear shifting and engine on / o ff ;similarly, the algebraic constraint h is defined as in (4), assum-ing T t is a known nonlinear function of v and P m is a knownnonlinear function of v and T m . We wish to minimize the totalpowertrain energy g ( x , u ) = γ f P f ( v , T e ) + γ q P q ( v , T m ), where P q = f ( x , u ) − E q and P f is a nonlinear mapping from the en-gine speed and torque to the fuel thermal power; the mappingfrom v to the engine speed is implicitly embedded.The input constraint set U ( v ) defines the speed-dependentactuator limits. The battery state of charge is confined to asafe operating region X . Its terminal value must exceed a pre-defined value, X N = X∩ (cid:110) E q : E (cid:63) q ≤ E q (cid:111) ; E (cid:63) q a ff ects the requiredcharging time after the trip, and therefore the waiting time untilthe vehicle is available for another trip.A technique to solve problem (6) was presented in [115]. Weapplied that technique to compute the optimal trajectory of thebattery charge for a typical commute in the Bay Area; the driv-ing data were measured on our plug-in hybrid electric test ve-hicle. Figure 6 compares the optimal charge trajectory with themeasured one. While the measured trajectory exhibits the typ-ical charge-depleting / charge-sustaining pattern, it can be seenthat the optimal strategy is to blend motor and engine usage.We refer to [115] for further analysis and implementation de-tails. Eco-driving often refers to the computation of a minimum-energy vehicle trajectory from an origin to a destination. Eco-driving exploits route information and long-term forecasts (likeroad grade and tra ffi c congestion) and accounts for constraintslike trip time and maximum velocity; vehicle stops and intersec-tions are also considered on urban and arterial roads. Here wefocus on scenarios in which no cooperating vehicles are avail-able.14 .5 1 1.5 2time (h)050100 b a tt e r y s t a t e o f c h a r g e ( % ) measuredoptimal Figure 6: Battery charge trajectories for a typical Bay Area commute: compar-ison between recorded data and optimal trajectory.
Literature review
While minimum fuel problems are classical in optimal con-trol [117, 175], the problem described above has some speci-ficity that we will clarify next, and is often denoted as (optimal) eco-driving problem. Below we summarize the eco-driving lit-erature for generic cruising scenarios and for the case of corri-dors of signalized intersections.
Reference cruising velocity generation.
When the CAV is totravel along a specified route, a reference velocity trajectory forthe on-board longitudinal control can be generated using routeinformation (like road topology, grade, curvature, and speedlimits) and dynamic data (like tra ffi c speed and weather fore-casts).In [176], the authors optimize the reference velocity for agiven route, considering road geometry, grade, tra ffi c informa-tion, and an accurate vehicle and powertrain model. Experi-ments show a fuel economy improvement between 5 and 15 %,when the problem is solved by dynamic programming in thecloud, and the reference velocity is tracked by a human driver.A similar problem, with a formulation in the time domain, isconsidered in [58]; by introducing some model simplifications,the optimal control policy is derived analytically. Numericalsolutions are also discussed for more general modeling assump-tions; the optimization methods include dynamic programmingand parametric optimization inspired by the analytical solution. Signalized intersections corridors.
Heuristic algorithms havebeen proposed to minimize braking and stopping at red lightsin [177, 118, 178, 179]. Several optimization-based algorithmshave also been proposed. The optimization goal may be to min-imize travel time, reduce acceleration peaks, idling at red lights,or directly minimize energy consumption. Dynamic program-ming is used in [180, 181, 182], while Dijkstra’s shortest pathalgorithms is used in [183, 184], MPC in [185, 186, 187, 188]and a genetic algorithm in [189]; in [190], the authors derive ananalytical solution for minimum energy driving through a cor-ridor of 3 intersections. [191] is, to the best of our knowledge,the only work reporting experimental results, in the case of aspeed advisory implementation. In practice, the problem being solved is a ff ected by uncer-tainty, due to tra ffi c, vehicle queues and pedestrians; further-more, in many cases intersections are adaptive to the tra ffi clevel, i.e. the phases duration is not fixed. Few of the citedworks explicitly consider these sources of uncertainty. A plan-ning method using a probabilistic signal timing forecast hasbeen proposed in [180]. In [192, 191, 193] the sensitivity of per-formance to a variety of factors linked to uncertainty (like con-gestion, penetration, communication range) is discussed. Therecent work [194] addresses the signal timing uncertainty ineco-driving systematically, by formulating and solving a robustoptimization problem; their formulation makes use of proba-bilistic or historical data of the signal timing. Challenges and opportunities for CAVs
As highlighted in [58], thus far the eco-driving concept hasbeen experimentally demonstrated as an extension of cruisecontrol systems.In optimal control formulations, tra ffi c speed is easily in-cluded as an upper bound on the vehicle speed; however, its un-certainty is generally neglected, with e ff ects that have not beeninvestigated thus far. Stop signs can also be included as stateconstraints in optimal control formulations; since they enforcea full vehicle stop, this approach essentially generates a multi-phase problem, which is acceptable if the travel time constraintis not tight. In the opposite case, in principle and if data areavailable, the intersection delay may be considered, as happensfor signalized intersections.When signalized intersection are included in the formulation,the open issues are multiple. A rigorous stochastic optimizationformulation for intersections with actuated signals has only re-cently been proposed [194]. If the assumption of free flow onthe road link is removed, forecasts of the tra ffi c state (vehicleoccupancy and speed) are required. In electric and hybrid pow-ertrains, avoiding vehicle stops may not always be the best pol-icy: the combination of regenerative braking and engine on / o ff may a ff ect significantly the optimal strategy. Example: eco-driving using signal timing data
We present a formulation of the optimal eco-driving prob-lem in the presence of signalized intersections, that was recentlyproposed in [194]. The problem is cast in the longitudinal posi-tion domain, as a finite-horizon optimal control problem of thefollowing form.minimize u , u ,..., u N − N − (cid:88) k = g ( x k , u k )subject to x k + = f ( x k , u k ) , u k ∈ U , x k ∈ X , k = , , . . . , N − , x = x s , x N ∈ X N . We set the state vector as x = [ t , v ] T and the input vector as u = [ F w , F b ] T , where t is the travel time, v is the vehicle speed, F w is the wheel force, F b is the braking torque. We model the15ongitudinal dynamics as in [39], project the time domain dy-namics into the position domain by the transformation dvdt = v dvds , and apply Euler discretization with step S s , obtaining f ( x , u ) = t + S s v v + S s Mv (cid:16) F w − F b − F f (cid:17) , where M is the vehicle mass and F f = Mg sin ϑ − Mg ( C r + C v v ) − ρ AC x v , g is the gravity constant, ϑ is the (position-dependent) roadslope, C r is the rolling coe ffi cient, C v is the viscous frictioncoe ffi cient, ρ is the air density A is the front area, C x is the airdrag coe ffi cient. Assuming a fuel-powered vehicle, the stagecost is set as the fuel rate g ( x , u ) = ˙ m f ( v , F w ) / v .The convex input constraint set U defines the actuator lim-its, while the state constraint set describes the surrounding en-vironment. A convex set X models bounds on the speed andthe acceleration. The formulation above can accommodate N s signalized intersections, assuming they can be approximatedas points along the route. We assume that every tra ffi c sig-nal has an independent cycle time c i ∈ (cid:104) , c i (cid:105) , i = , . . . , N s ,where c i = c i ∈ R + is the cycle period. We denote the red light phaseduration by c ir ∈ (cid:16) , c i (cid:17) , and by t ip the time at which the CAVpasses through intersection i . In the domain of the intersectioncycle time c i , the passing time is computed as c ip = (cid:16) c i + t ip (cid:17) mod c i , where c i is the cycle time at s =
0. We enforcethat intersections are not crossed during red light phases by X = (cid:110) t : c ip ( t ) ≥ c ir , ∀ i = , . . . , N s (cid:111) . In sum, the state con-straint set is given by X = X ∩ X .Thus far, exact forecasts of the red light phase durations c ir were assumed available throughout the route. In practice, manyintersections adapt their phase durations based on the time ofthe day and on the tra ffi c level, making perfect forecasts unreal-istic. In this case, [194] proposed to replace X with the chanceconstraint X = (cid:110) t : Pr (cid:104) c ip ( t ) ≥ c ir + α i (cid:105) ≥ − η i , ∀ i = , . . . , N s (cid:111) , where Pr [ A ] is the probability of event A , α i ∈ (cid:104) , c i − c ir (cid:105) mod-els the adaptation of the red light phase, and η i ∈ [0 ,
1] is thelevel of constraint enforcement.Figure 7 shows the solutions to the deterministic problem(i.e. enforcing X = X ∩ X ) and to some instances of therobust problem (i.e. enforcing X = X ∩ X for di ff erent valuesof η i ). It can be noted that the deterministic solution crossesthe third intersection very close to a phase switching (red togreen); conversely, the robust solutions show di ff erent levels ofconservatism, which can by adjusted by tuning η i . We referto [194] for an extensive analysis, implementation details, andapproaches to define X based on historical signal timing data. Figure 7: Eco-driving through signalized intersections.
In multi-vehicle CAV applications, the characterizing featureis vehicle cooperation: the approaches presented here assumecommunication with other vehicles via V2V communication,with a road-side coordinator via V2I communication, with aremote, cloud-based coordinator via cellular communication,or a combination thereof. The optimization problems involvegroups of vehicles of variable size; in this sense, these applica-tions are at the border of tra ffi c control, that - roughly speaking- tackles similar problems at the road network level, rather thanthe vehicle level. For a survey of tra ffi c control and its links tovehicle connectivity we refer to [195, 196]. Literature review
The literature on multi-vehicle trajectory planning is ex-tremely vast. Multi-vehicle coordination and planning havebeen thoroughly studied for autonomous robots, unmannedaerial vehicles, marine vehicles. Here we survey some multi-vehicle coordination problems that arise in CAVs, i.e. coordina-tion on autonomous roadways (that includes speed harmoniza-tion and coordination at merging roadways and autonomous in-tersections) and platoon coordination . A pictorial classificationof these applications is given in Figure 8.
Coordination on autonomous roadways.
A problem of multi-vehicle coordination that arises in automated highways is speedharmonization . Speed harmonization consists in controlling thespeed of vehicles before they reach a speed reduction zone (Fig-ure 8). Speed reduction zones are congested because of roadworks, tollbooths, or accidents. In today’s highways, speed har-monization is implemented using variable speed limits or patrolvehicles [197].In a fully automated setting, it is possible to control the speedtrajectories of the individual CAVs [198]. If CAVs are not al-lowed to change lanes, the only safety requirement is to avoidrear end collisions. If they are also assumed to perfectly tracktheir optimal trajectories, the problem can be solved in a fullydecentralized manner, every time that a new CAV enters the16 a) Speed harmonization. (b) Merging roadways coordination. (c) Autonomous intersection. (d) Platoon coordination.Figure 8: Coordination problems for groups of Connected and Automated Vehicles (CAVs). CAVs are represented in white (Created on https://icograms . com ). control zone. An analytical solution to the speed harmoniza-tion problem is given in [198, 199]. Using a microscopic tra ffi csimulator, the optimal control of CAVs is compared to human-driven vehicles and to speed harmonization via variable speedlimits and patrol vehicles.The problem of smoothly merging or intersecting twostreams of vehicles, without provoking stop-and-go driving, isreminiscent of the speed harmonization problem. The main dif-ference is that safety guarantees require to avoid not only rear-end collisions, but also lateral collisions (at the point where thetwo streams merge or cross). Two prototypical instances of thisproblem are merging roadways and autonomous intersections (Figure 8). In these two scenarios, the tra ffi c flow is currentlyregulated by ramp meters and tra ffi c lights; both problems havebeen extensively studied by the tra ffi c control community.Recent research has revisited these two problems under theassumption that all the vehicles on the road are CAVs. Theseproblems have raised great interest in the control community,which has produced a vast literature on the topic. Due to spacelimitations, we refer the interested reader to the recent survey[200], that has summarized these e ff orts. Even limiting thescope to optimization-based approaches, the possible problemformulations are multiple; formulations di ff er for the optimiza-tion objective (including travel time, road throughput, fuel con-sumption, or combinations thereof) and for the way safety con-straints are enforced. A simplified problem can be formulatedassuming that [201] (i) vehicles entering the control zone areserved on a FIFO basis, (ii) only one vehicle at a time is allowedin the merging zone, (iii) no turns are allowed in the mergingzone, and (iv) vehicles cross the merging zone at constant veloc-ity. The main di ff erence with the speed harmonization problemis that multiple vehicles are considered jointly, and that onlyone vehicle at a time is allowed in the merging zone to avoidlateral collisions. The exit times from the merging zone are,in principle, free optimization variables. Because a FIFO pol-icy is adopted in the control zone, the exit time of each vehicleis then upper bounded by that of the preceding vehicle, giventhe collision avoidance constraints. Using this argument, theproblem can be divided into decentralized problems, similar tothe speed harmonization problem, although with some risk ofconservatism. Platoon coordination.
While the algorithms in the previousparagraph generally assume that all the agents on the road areCAVs, in this paragraph this is only assumed for the agents inthe platoon. As discussed in Section 4.2, longitudinal control ofplatoons is mostly regarded as a distributed control problem, butsome basic level of centralized coordination is still required; inparticular, the formation geometry (inter-vehicular spacing pol-icy, cruising speed, vehicle ordering) and the information flownetwork (communication topology and quality, communicationprotocol) a ff ect platoon safety and performance. These aspectshave not been standardized to date, and due to the heterogeneityof vehicles, sensors, actuators, and communication technolo-gies available on the market, they could reasonably be coor-dinated remotely on a case-by-case basis. For instance, somecommunication topologies listed in Figure 4 may not be feasi-ble, depending on the number and the specific instrumentationof the vehicles involved in the platoon. Another example is thechoice of the inter-vehicle spacing policy, which depends on anumber of factors, including the dynamics and non-idealities ofsensors, actuators, and wireless communication [202, 203].For a formed platoon, motion is mostly longitudinal withina lane, although lateral motion is needed for lane changesand collision avoidance. Platoon formation and dismantlingis a closely related problem which also requires coordination.Highway platoon formation and management is discussed in[204], where the conjecture is that highway platoons should re-main intact for as long as possible. The paper develops andanalyzes strategies to sort vehicles and form platoons at thehighway entrances, in order to maximize the distance that thevehicles can travel together, so that the platoon does not needreorganization. A CAV that joins or leaves a platoon needs au-thority on both longitudinal and lateral motion, while the ve-hicles in the platoon may only move longitudinally to open orclose a gap. Merging of (or splitting into) two platoons is notfundamentally di ff erent; instead of just one gap, there may bemultiple gaps to open or close. Merging is a critical maneu-ver because the trail vehicle or platoon must temporarily movefaster than the lead platoon, and with a smaller distance gap: asudden deceleration of the lead platoon may result in a colli-sion at high speed. Thus, their relative velocities must be keptbounded to avoid dangerous collisions. An approach for pla-17oon merging, splitting, and lane changing is presented in [205].The authors determine a maximum safe velocity for the trailplatoon, as a function of spacing and lead-platoon velocity; theproposed merge, split and lane changing maneuvers keep thevelocity of the trail platoon below that limit. [206] is focusedon a merging protocol, with particular focus on the communi-cations exchanged between vehicles, for a similar scenario.A problem at a slightly higher level is the clustering of ve-hicles into platoons, based on their routes and departure andarrival times. [127] presents a hierarchical control architec-ture for freight transportation, including platoon merging, split-ting or reordering. Speed trajectories to merge into a grow-ing platoon are computed in [207] using an optimal control ap-proach. The algorithm receives the origins, destinations, andtimes of departure and arrival of the lead platoon and of thevehicles joining along the road; using a hybrid systems exten-sion of Pontryagin’s minimum principle, it computes the opti-mal merging times and the corresponding velocity trajectories.Another opportunity lies in the coordination of vehicle fleets:when origins, destinations and time constraints are known inadvance, a centralized planner can be set up, aggregating vehi-cles in platoons and thereby maximizing the overall fuel econ-omy [208, 209]. In [210], a multi-agent system approach is pro-posed for the management of an autonomous intersection,whereCAVs may form platoons. Simulations show improvementscompared to traditional intersection control; when comparedto a non-platoon based autonomous intersection, the commu-nication load is shown to decrease substantially, and the sys-tem appears more robust against tra ffi c volume variations. In[185, 186] the authors propose a MPC framework for groups ofCAVs driving through an urban corridor, including signalizedintersections. A decentralized approach is taken, as every CAVonly receives information from neighboring vehicles and tra ffi csignals. Challenges and opportunities for CAVs
In the presented multi-vehicle coordination problems, a uni-fying aspect is that, once the vehicles are engaged, the maneu-vers often become safety critical and rely heavily on the on-board algorithms. For example, for a CAV approaching a pla-toon, the remote coordinator may only send high-level indica-tions (where in the platoon to open a gap, the dimension of thegap, etc.). For deployment on public roads, planning will re-quire updating in real-time to react to the surrounding tra ffi c.For all these reasons, the real-time coordination between theremote planner and the on-board controls is critical. An oppor-tunity in this sense is the implementation of these algorithms,to the extent possible, in a distributed manner.The research on autonomous roadways has demonstratedhigh potential in many scenarios. However, further research isneeded to deploy these solutions on public roads, having CAVsinteract with other vehicles. Open questions include both thereformulation of the coordination problems, and the e ff ect onperformance of partial penetration of CAVs.In platoon coordination, a framework that unifies di ff er-ent communication topologies is lacking. Further research isrequired to balance safety and robustness requirements (like string stability) with broader impacts (like energy consumptionand road throughput). Example: MPC for platoon coordination
We consider the following platoon coordination problem:given the origins and destinations of a group of V agents travel-ing on the same route, compute the optimal longitudinal trajec-tories, allowing changes of vehicle order. Aside from collisionavoidance, the formation geometry is free: one, more, or noplatoon can be formed, as long as the origin and destinationconstraints are satisfied. We formulate the problem in the timedomain, as a finite-horizon optimal control problem of the fol-lowing form.minimize u i , u i ,..., u iNi − V (cid:88) i = N i − (cid:88) k = g i ( x k , u k )subject to x ik + = f i ( x ik , u ik ) , = h i ( x ik , u ik ) , u ik ∈ U i , x ik ∈ X i , k = , . . . , N i − , i = , . . . , V , x i = x it , x iN i ∈ X iN i , i = , . . . , V . (7)The index i is a unique vehicle identifier. For each vehicle i ,we set the state vector as x i = [ s i , v i , p i ] T and the input vectoras u i = [ F iw , F ib , u ip ] T , where d i is the longitudinal position, v i is the longitudinal speed, p i ∈ { , . . . , V } is the position in theplatoon (from first to last), F iw is the wheel force, F ib is the brak-ing torque, u ip ∈ {− , , + } is a discrete variable that initiatesa change of position with the preceding or following vehicle.We model the longitudinal dynamics as in [39] and apply Eulerdiscretization with step T s , obtaining f ( x , u ) = s i + T s v i v i + T s M (cid:16) F iw − F ib − F if (cid:17) p i + u ip , where M is the vehicle mass and F if = Mg sin ϑ − Mg ( C r + C v v i ) − ρ AC x ( d i )( v i ) , g is the gravity constant, ϑ is the (position-dependent) roadslope, C r is the rolling coe ffi cient, C v is the viscous frictioncoe ffi cient, ρ is the air density A is the front area, C x is the airdrag coe ffi cient, that is a non-increasing function of the inter-vehicular distance. All the parameters can di ff er from vehi-cle to vehicle (we neglected the index i for simplicity). Theconstraint h models the (formation-dependent) computation ofinter-vehicular distances, d i = s p i + − s p i − L i ∈ R + . We wish tominimize the total energy at the wheel g i ( x i , u i ) = F iw v i .The input constraint set U i defines the actuator limits; thisincludes enforcing ( u ip = + → ( u i + p = −
1) and ( u ip = − → ( u i − p = + igure 9: Velocity trajectories of 3 vehicle under the platoon management strat-egy based on distributed MPC proposed in [211]. approximation, the problem is complex due to the mixed inte-ger nonlinear dynamics and the dimension of the state space. Acomputationally tractable approach to approximate the optimalsolution to problem (7) haw been proposed in [211], where areceding horizon approximation is taken, smooth dynamics areused, and the problem is solved in a distributed way. Figure 9shows the position trajectories for a group of three vehicles withdi ff erent origins and destinations. Although the vehicles havedi ff erent origins and destinations, and the problem is solved ina distributed receding horizon fashion, it can be noted that thethree vehicles form a platoon in the central part of the trip. Werefer to [211] for a detailed performance analysis. Classical algorithms for vehicle routing search the shortest(minimum distance) or fastest (minimum time) route from anorigin to a destination [117]. Eco-routing pursues the route onwhich the vehicle incurs minimum energy consumption.
Literature review
Given a deterministic and time invariant model of energyconsumption, the energy-optimal routing is simply a shortestpath problem. Di ff erent model structures have been used, in-cluding the Comprehensive Modal Emissions Model [212, 213,214], data-driven models [215], and physical models based onthe vehicle longitudinal dynamics [216, 217]. Other works haveintroduced time-varying and uncertain models [215, 218, 219].An objective comparison of the routing methods proposed in[212, 214, 215, 216] is presented in [220]; the methods di ff er inhow they compute the edge costs, either directly from data (likeGPS traces) or using vehicle models (like the standard longi-tudinal model). The di ff erent eco-routes are compared to theshortest and fastest routes, for a large set of origins and des-tinations, using SUMO [221] to generate realistic tra ffi c pat-terns. The energy consumption model is found to be critical forperformance: in some cases, the eco-routes lead higher energyconsumption than the fastest route. Another limitation of eco-routes is that they may turn out rel-atively time consuming or lengthy [220]. To address this issue,one may resort to multi-objective and constrained shortest pathalgorithms [117]. In the first case, the algorithm returns a Paretooptimal route, that balances fuel consumption, travel time anddistance. In the second case, the minimization of fuel consump-tion is subject to constraints on the maximum travel time and / oron the maximum travel distance.Some eco-routing methods are specifically tailored for elec-trified powertrains: given the limited amount of energy that canbe stored in a battery, the eco-routing problem is even morecompelling. To give a guarantee on the driving range, the algo-rithms must keep track of the energy stored in the battery, whichadds complexity to the problem. The possibility to perform re-generative braking leads to negative energy cost in some roadsegments, which requires to modify the standard routing algo-rithms. Recent works [216, 222] have addressed this problemalso for hybrid powertrains, which present an additional chal-lenge: their energy consumption generally includes both fueland grid electricity, and their usage along the route is defined inreal-time by the energy management system (see Section 4.1).In [222], a simplified energy management strategy is assumed.Finally, electric and plug-in hybrid vehicles can recharge theirbattery, hence stops at charging stations may be included. Thesame can be said for fuel stations, but the range issue is not aspressing and refuelling is much faster than battery charging.The problem of minimum time routing with limited en-ergy and including stops at charging stations is studied in[223, 224, 225]. The problem is solved using mixed integernon-linear programming in [223, 224] and dynamic program-ming in [225]. A multi-vehicle extension is also studied in[225], including tra ffi c congestion e ff ects; to mitigate the com-putational complexity in this scenario, an alternative flow opti-mization formulation is proposed.In [226], the eco-routing problem is studied for a signalizedtra ffi c network. This brings additional complexity into the prob-lem, because the estimation of velocity trajectories over links(usually approached using historical data) becomes even morechallenging and uncertain. The tra ffi c network is modeled as aMarkov decision process, and the edge costs are estimated witha microscopic vehicle emission model (as in [212, 214]). Challenges and opportunities for CAVs
Several aspects of eco-routing deserve further investigation.As we discussed, common pitfalls are model accuracy anduncertainty. The application of eco-routing to CAVs seemspromising in this regard: the on-board controls, removing tosome extent the human driver from the loop, lead to more con-sistent energy consumption. Another direction that has beenlittle investigated is the use of systematic methods to handleuncertainty in models and forecasts. Finally, the e ff ect of eco-routing (and routing algorithms in general) at the network level(rather than at the vehicle level only) is not well understoodyet. A big challenge in this sense is that large scale deploymentof these technologies is generally prohibitive for academic re-searchers.19 xample: eco-routing for a plug-in hybrid electric vehicle Routing algorithms generally search a path in a graph G = ( N , E ), where the nodes in the set N = { n k : k = . . . N } rep-resent intersections and other important road locations, and theedges in the set E = { e k : k = . . . E } represent the road seg-ments connecting the nodes. A route or path is a sequence ofcontiguous nodes p = { n O , . . . , n D } , with n O and n D the originand destination nodes, respectively. A simple path is a pathwhere every node is visited at most once. P is the set of allsimple paths in the map G and P = | p | is cardinality of path p .We formulate the eco-routing (or minimum energy routing)problem as follows.minimize n , n ,..., n P , P P (cid:88) k = c ( n k , n k + , x k )subject to x k + = x k + E q ( n k , n k + , x k ) , x k ∈ X , k = , . . . , P , { n , n , . . . , n P } ∈ P , n = n O , n P = n D , x = x , x P ∈ X (cid:63) . (8)The minimum energy path (eco-route) p e is computed based onthe edge costs c ( n k , n k + , x k ) : N × N × R + → R + ; we take itas a weighted sum of the fuel energy E f and battery energy E q consumed along the edge e ( n k , n k + ) c ( n k , n k + , x k ) = γ f E f ( n k , n k + , x k ) + γ q E q ( n k , n k + , x k ) . Here E f and E q are function of the current and next node, n k and n k + , and of the current battery energy x k . In reality E f and E q are complex functions of the vehicle speed, the road gradeand curvature, the vehicle and powertrain dynamics, and theon-board control strategies. For instance, in a plug-in hybridelectric vehicle, the usage of fuel and battery energy is highlydependent on the state x k ; this is the first reason why x is in-cluded in the formulation. In practice, E f and E q are estimatedbased on vehicle and powertrain models, and on the availableroute data for the edge e ( n k , n k + ); route data can include grade,curvature, speed limits, tra ffi c speed, weather.A second reason to include x in the formulation is to con-strain the battery charge. In the problem above, this is as simpleas enforcing a safe operating range X throughout the route, anda target charge X (cid:63) at destination. As mentioned previously, theterminal charge a ff ects the charging time, i.e. the down timeuntil the next trip. With minimal modifications, the formulationabove can accommodate stops at charging stations also alongthe route.Figure 10 shows a comparison between the minimum dis-tance route and the minimum energy route, for an origin and adestination in the Berkeley area. The minimum distance routeis 9 .
72 km long, and requires 4 .
23 kWh according to a simplemodel of plug-in hybrid electric vehicle. The minimum energyroute is 10 .
03 km long, and requires 3 .
22 kWh according to thesame model. A computational approach to solve problem (8)has been proposed in [222]. -122.34 -122.32 -122.3 -122.28 -122.26longitude [deg]37.8737.8837.8937.937.9137.92 l a tit ud e [ d e g ] shortest routeeco-route Figure 10: Minimum distance and minimum energy routes for an origin and adestination in the Berkeley area.
6. Conclusion and outlook
Driving automation and vehicle connectivity are progres-sively becoming part of our everyday after three decades ofresearch e ff orts. The role of control and planning is crucialto safety and performance: this paper surveyed the existingliterature on these algorithms at di ff erent hierarchical levels.Alongside with the main components and technologies for con-nected and automated vehicles, we identified a possible controland planning architecture. We framed the existing approacheswithin this architecture, with a twofold objective: examine thestate of the art on the various technologies, and identify theirrole in the bigger picture.This system level approach helped identify challenges anduntapped opportunities. Control and planning algorithms musthave a well defined scope; however, the interactions (and insome cases the integration) between di ff erent functional blockshave not been exhaustively investigated.For most technologies, we also noticed a lack of experimen-tal validation. While testing on public roads is still very chal-lenging, the current state of technology o ff ers the opportunityto deploy advanced algorithms on real vehicles. Selecting ap-propriate testing scenarios, that are representative of real-worldconditions, is a non-trivial open question. The current direc-tions pursued by public authorities, private companies and aca-demic researchers seem to support this view, and a field valida-tion of the topics presented in this paper can be expected in thenear future. Acknowledgement
The information, data, or work presented herein was fundedin part by the Advanced Research Projects Agency-Energy(ARPA-E), U.S. Department of Energy, under Award Numberde-ar0000791. The views and opinions of authors expressed20erein do not necessarily state or reflect those of the UnitedStates Government or any agency thereof.Figure 7 has been extracted from [194] and generated by Dr.Chao Sun, to whom the authors are grateful.
References [1] S. E. Shladover, C. A. Desoer, J. K. Hedrick, M. Tomizuka, J. Walrand,W. B. Zhang, D. H. McMahon, H. Peng, S. Sheikholeslam, N. McKe-own, Automatic Vehicle Control Developments in the PATH Program,IEEE Transactions on Vehicular Technology 40 (1) (1991) 114–130.[2] S. E. Shladover, PATH at 20 – History and Major Milestones, 2006 IEEEIntelligent Transportation Systems Conference (1) (2006) 22–1.[3] E. D. Dickmanns, V. Graefe, Dynamic monocular machine vision andapplications of dynamic monocular vision, Machine Vision and Appli-cations 1 (1988) 223–240.[4] E. D. Dickmanns, Vehicles Capable of Dynamic Vision, in: 15th Inter-national Joint Conference on Artifical Intelligence, IJCAI’97, MorganKaufmann Publishers Inc., San Francisco, CA, USA, 1997, pp. 1577–1592.[5] CMU NAVLAB - No Hands Across America Home Page.URL . cs . cmu . edu/~tjochem/nhaa/nhaa home page . html [6] M. Buehler, K. Iagnemma, S. Singh, The 2005 DARPA Grand Chal-lenge, Springer, New York, 2007.[7] M. Buehler, K. Iagnemma, S. Singh, The DARPA Urban Challenge,Springer, New York, 2009.[8] J. Xin, C. Wang, Z. Zhang, N. Zheng, China future challenge: Beyondthe intelligent vehicle, IEEE Intell. Transp. Syst. Soc. Newslett. 16 (2)(2014) 8–10.[9] P. Cerri, G. Soprani, P. Zani, J. Choi, J. Lee, D. Kim, K. Yi, A. Broggi,Computer vision at the Hyundai Autonomous Challenge, in: IEEE Con-ference on Intelligent Transportation Systems, 2011, pp. 777–783.[10] A. Broggi, P. Cerri, M. Felisa, M. C. Laghi, L. Mazzei, P. P. Porta, TheVisLab Intercontinental Autonomous Challenge: an extensive test fora platoon of intelligent vehicles, International Journal of Vehicle Au-tonomous Systems 10 (3) (2012) 147.[11] A. Broggi, P. Cerri, S. Debattisti, M. C. Laghi, P. Medici, D. Molinari,M. Panciroli, A. Prioletti, PROUD-Public Road Urban Driverless-CarTest, IEEE Transactions on Intelligent Transportation Systems 16 (6)(2015) 3508–3519.[12] J. Ziegler, P. Bender, M. Schreiber, H. Lategahn, T. Strauss, C. Stiller,T. Dang, U. Franke, N. Appenrodt, C. G. Keller, E. Kaus, R. G. Her-rtwich, C. Rabe, D. Pfei ff er, F. Lindner, F. Stein, F. Erbs, M. En-zweiler, C. Knoppel, J. Hipp, M. Haueis, M. Trepte, C. Brenk, A. Tamke,M. Ghanaat, M. Braun, A. Joos, H. Fritz, H. Mock, M. Hein, E. Zeeb,Making bertha drive-an autonomous journey on a historic route, IEEEIntelligent Transportation Systems Magazine 6 (2) (2014) 8–20.[13] Autopilot — Tesla.URL . tesla . com/autopilot [14] Ford Autonomous 2021.URL https://corporate . ford . com/innovation/autonomous-2021 . html [15] Cruise Automation.URL https://getcruise . com/ [16] Self-driving car technology — Bosch Global.URL . bosch . com/explore-and-experience/autonomous-driving-interview-with-moritz-dechant/ [17] nuTonomy - Home.URL . nutonomy . com/ [18] Waymo.URL https://waymo . com/ [19] Self-Driving Cars — Uber.URL . uber . com/cities/pittsburgh/self-driving-ubers/ [20] Taxonomy and Definitions for Terms Related to On-Road Motor VehicleAutomated Driving Systems (2014).[21] S. Kato, S. Tsugawa, K. Tokuda, T. Matsui, H. Fujii, Vehicle Control Al-gorithms for Cooperative Driving with Automated Vehicles and Interve- hicle Communications, IEEE Transactions on Intelligent TransportationSystems 3 (3) (2002) 155–160.[22] E. van Nunen, M. R. J. a. E. Kwakkernaat, J. Ploeg, B. D. Netten, Co-operative Competition for Future Mobility, IEEE Transactions on Intel-ligent Transportation Systems 13 (3) (2012) 1018–1025.[23] C. Englund, L. Chen, J. Ploeg, E. Semsar-Kazerooni, A. Voronov,H. H. Bengtsson, J. Dido ff , The Grand Cooperative Driving Challenge2016: Boosting the introduction of cooperative automated vehicles,IEEE Wireless Communications 23 (4) (2016) 146–152.[24] T. Robinson, E. Chan, E. Coelingh, Operating Platoons On Public Mo-torways : An Introduction To The SARTRE Platooning Programme, in:17th ITS World Congress on Intelligent Transportation Systems, 2010.[25] S. Tsugawa, S. Kato, K. Aoki, An automated truck platoon for energysaving, in: IEEE International Conference on Intelligent Robots andSystems, 2011, pp. 4109–4114.[26] V2V Safety Technology Now Standard on Cadillac CTS Sedans (2017).URL http://media . cadillac . com/media/us/en/cadillac/news . detail . html/content/Pages/news/us/en/2017/mar/0309-v2v . html [27] Audi announces the first vehicle to infrastructure (V2I) service - thenew Tra ffi c light information system — Audi USA.URL . audiusa . com/newsroom/news/press-releases/2016/08/audi-announces-first-vehicle-to-infrastructure-service [28] Dedicated Short Range Communications (DSRC) Message Set Dictio-nary (2016).[29] W. Knight, Car-to-Car Communication - MIT Technology Review(2015).URL . technologyreview . com/s/534981/car-to-car-communication/ [30] J. H. Gawron, G. A. Keoleian, R. De Kleine, T. J. Wallington, H. C. Kim,Life Cycle Assessment of Connected and Automated Vehicles (CAVs):Sensing and Computing Subsystem and Vehicle Level E ff ects, Environ-mental science & technology.[31] S. Sill, DSRC: The Future of Safer Driving.URL . its . dot . gov/factsheets/dsrc factsheet . htm [32] ETSI TR 101 607 V1.1.1 (2013).[33] Savari, Inc.URL savari . net [34] Arada Systems.URL . aradasystems . com/ [35] DENSO ADAS.URL . denso . com/us-ca/en/innovation/technology/adas/ [36] Cohda Wireless.URL . cohdawireless . com/ [37] N. Alam, A. Tabatabaei Balaei, A. G. Dempster, A DSRC Doppler-Based Cooperative Positioning Enhancement for Vehicular NetworksWith GPS Availability, IEEE Transactions on Vehicular Technology60 (9) (2011) 4462–4470.URL http://ieeexplore . ieee . org/document/6020821/ [38] A. Nordrum, Autonomous Driving Experts Weigh 5G Cellular Net-work Against Dedicated Short Range Communications, IEEE SpectrumMay (Cars That Think).[39] L. Guzzella, A. Sciarretta, Vehicle Propulsion Systems, Springer, 2013.[40] D. Gibson, M. K. P. Mills, L. A. Klein, A New Look at Sensors, PublicRoads.URL . fhwa . dot . gov/publications/publicroads/07nov/04 . cfm [41] Caltrans Performance Measurement System (PeMS).URL http://pems . dot . ca . gov/ [42] J. Lioris, R. Pedarsani, F. Y. Tascikaraoglu, P. Varaiya, Platoons of con-nected vehicles can double throughput in urban roads, TransportationResearch Part C: Emerging Technologies 77 (2017) 292–305.[43] A. Muralidharan, C. Flores, P. Varaiya, High-resolution sensing of urbantra ffi c, in: 17th IEEE Conference on Intelligent Transportation Systems,2014, p. 780.[44] S. Calvert, H. Mahmassani, J.-n. Meier, P. Varaiya, S. Hamdar, D. Chen,X. Li, A. Talebpour, S. P. Mattingly, . C. A. V. . Deployment, . C. .Urban, Tra ffi c Flow of Connected and Automated Vehicles: Challenges nd Opportunities, in: Road Vehicle Automation 4, 2018, pp. 235–245.[45] C. M. Day, D. M. Bullock, H. Li, S. M. Remias, A. M. Hainen, R. S.Freije, A. L. Stevens, J. R. Sturdevant, T. M. B. Jr., Performance Mea-sures for Tra ffi c Signal Systems: An Outcome-Oriented Approach,Tech. rep. (2014).[46] A. Muralidharan, S. Coogan, C. Flores, P. Varaiya, Management of inter-sections with multi-modal high-resolution data, Transportation ResearchPart C: Emerging Technologies 68 (2016) 101–112.[47] K. Clement-Nyns, E. Haesen, J. Driesen, The impact of Charging plug-in hybrid electric vehicles on a residential distribution grid, IEEE Trans-actions on Power Systems 25 (1) (2010) 371–380.[48] K. Clement-Nyns, E. Haesen, J. Driesen, The impact of vehicle-to-gridon the distribution grid, Electric Power Systems Research 81 (1) (2011)185–192.[49] Z. Fan, A distributed demand response algorithm and its application toPHEV charging in smart grids, IEEE Transactions on Smart Grid 3 (3)(2012) 1280–1290.[50] Google Maps API.URL https://developers . google . com/maps/ [51] HERE.URL . here . com/en [52] Inrix.URL http://inrix . com/ [53] S. Ibrahim, D. Kalathil, R. O. Sanchez, P. Varaiya, Estimating PhaseDuration for SPaT Messages (2017).URL http://arxiv . org/abs/1710 . [54] Sensis Networks.URL . sensysnetworks . com/ [55] J. van de Sluis, O. Baijer, L. Chen, H. H. Bengtsson, L. Garcia-Sol,P. Balaguer, Proposal for extended message set for supervised automateddriving, Tech. rep., TNO, VIKTORIA, IDIADA (2015).[56] Peloton Technology.URL https://peloton-tech . com/ [57] J. M. Owens, R. Greene-roesel, A. Habibovic, L. Head, Reducing Con-flict Between Vulnerable Road Users and Automated Vehicles, in: RoadVehicle Automation 4, 2018, pp. 69–75.[58] A. Sciarretta, G. De Nunzio, L. L. Ojeda, Optimal Ecodriving Control:Energy-E ffi cient Driving of Road Vehicles as an Optimal Control Prob-lem, IEEE Control Systems Magazine 35 (5) (2015) 71–90.[59] B. Paden, M. Cap, S. Z. Yong, D. Yershov, E. Frazzoli, A Survey of Mo-tion Planning and Control Techniques for Self-driving Urban Vehicles,IEEE Transactions on Intelligent Vehicles 1 (1) (2016) 33–55.URL http://arxiv . org/abs/1604 . [60] A. Al Alam, B. Besselink, V. Turri, J. Martensson, K. H. Johansson,Heavy-duty vehicle platooning for sustainable freight transportation: Acooperative method to enhance safety and e ffi ciency, IEEE Control Sys-tems Magazine 35 (6) (2015) 34–56.[61] A. Carvalho, S. Lef´evre, G. Schildbach, J. Kong, F. Borrelli, Automateddriving: The role of forecasts and uncertainty - A control perspective,European Journal of Control 24 (2015) 14–32.[62] V. Ngo, T. Hofman, M. Steinbuch, A. Serrarens, Optimal Control of thegearshift command for hybrid electric vehicles, IEEE Transactions onVehicular Technology 61 (8) (2012) 3531–3543.[63] T. N¨uesch, A. Cerofolini, G. Mancini, N. Cavina, C. Onder, L. Guzzella,Equivalent consumption minimization strategy for the control of realdriving NOx emissions of a diesel hybrid electric vehicle, Vol. 7, 2014.[64] M. Josevski, D. Abel, Gear shifting and engine on / o ff optimal controlin hybrid electric vehicles using partial outer convexification, in: IEEEConference on Control Applications, 2016, pp. 562–568.[65] B. Saerens, M. Diehl, E. Van Den Bulck, Optimal Control Using Pon-tryagin’s Maximum Principle and Dynamic Programming, in: L. delRe, F. Allgower, L. Glielmo, C. Guardiola, I. V. Kolmanovsky (Eds.),Automotive Model Predictive Control. Lecture Notes in Control and In-formation Sciences, Vol. 402, Springer, London, 2010.[66] T. J. Boehme, B. Frank, M. Schori, T. Jeinsch, Multi-Objective Opti-mal Design of Parallel Plug-In Hybrid Powertrain Configurations withRespect to Fuel Consumption and Driving Performance, in: 2014 Euro-pean Control Conference, 2014, pp. 1017–1023.[67] L. Guzzella, C. H. Onder, Introduction to modeling and control of inter-nal combustion engine systems, 2010.[68] J. Bishop, A. Nedungadi, G. Ostrowski, B. Surampudi, P. Armiroli, E. Taspinar, An Engine Start / Stop System for Improved Fuel Economy,in: 2007 SAE World Congress, 2007.URL http://papers . sae . org/2007-01-1777/ [69] H. Koch-groeber, J. Wang, Criteria for Coasting on Highways for Pas-senger Cars, SAE Technical Paper 01 (1157).[70] A. Balluchi, M. Di Benedetto, C. Pinello, C. Rossi, A. Sangiovanni-Vincentelli, Cut-o ff in Engine Control: a Hybrid System Approach, in:36th Conference on Decision and Control, no. December, 1997, pp.4720–4725.[71] M. Canova, K. Sevel, Y. Guezennec, S. Yurkovich, Control of theStart / Stop of a Diesel Engine in a Parallel HEV with a BeltedStarter / Alternator, in: 8th International Conference on Engines for Au-tomobile, 2007.[72] M. Canova, Y. Guezennec, S. Yurkovich, On the Control of EngineStart / Stop Dynamics in a Hybrid Electric Vehicle, Journal of DynamicSystems, Measurement, and Control 131 (6) (2009) 061005.[73] P. Elbert, T. Nuesch, A. Ritter, N. Murgovski, L. Guzzella, EngineOn / O ff control for the energy management of a serial hybrid electric busvia convex optimization, IEEE Transactions on Vehicular Technology63 (8) (2014) 3549–3559.[74] N. Murgovski, L. Johannesson, J. Sj¨oberg, Convex modeling of en-ergy bu ff ers in power control applications, IFAC Proceedings Volumes(IFAC-PapersOnline) (2012) 92–99.[75] F. R. Salmasi, Control Strategies for Hybrid Electric Vehicles: Evolu-tion, Classification, Comparison, and Future Trends, IEEE Transactionson Vehicular Technology 56 (5) (2007) 2393–2404.[76] S. G. Wirasingha, A. Emadi, Classification and review of control strate-gies for plug-in hybrid electric vehicles, IEEE Transactions on VehicularTechnology 60 (1) (2011) 111–122.[77] A. A. Malikopoulos, Supervisory power management control algorithmsfor hybrid electric vehicles: A survey, IEEE Transactions on IntelligentTransportation Systems 15 (5) (2014) 1869–1885.[78] P. Pisu, G. Rizzoni, A comparative study of supervisory control strate-gies for hybrid electric vehicles, IEEE Transactions on Control SystemsTechnology 15 (3) (2007) 506–518.[79] D. Maamria, F. Chaplais, N. Petit, A. Sciarretta, Comparison of sev-eral strategies for HEV energy management system including engine andcatalyst temperatures, Proceedings of the American Control Conference2015-July (2015) 2948–2955.[80] S. Delprat, J. Lauber, T.-M. Guerra, J. Rimaux, Control of a parallelhybrid powertrain: optimal control, Vehicular Technology, IEEE Trans-actions on 53 (3) (2004) 872–881.[81] S. Barsali, C. Miulli, a. Possenti, A Control Strategy to Minimize FuelConsumption of Series Hybrid Electric Vehicles, IEEE Transactions onEnergy Conversion 19 (1) (2004) 187–195.[82] J. S. Won, R. Langari, Intelligent energy management agent for a paral-lel hybrid vehicle - Part II: Torque distribution, charge sustenance strate-gies, and performance results, IEEE Transactions on Vehicular Technol-ogy 54 (3) (2005) 935–953.[83] J. S. Won, R. Langari, M. Ehsani, An energy management and chargesustaining strategy for a parallel hybrid vehicle with CVT, IEEE Trans-actions on Control Systems Technology 13 (2) (2005) 313–320.[84] A. Sciarretta, L. Guzzella, Control of Hybrid Electric Vehicles, IEEEControl Systems Magazine 27 (April) (2007) 60–70.[85] A. Sciarretta, M. Back, L. Guzzella, Optimal control of parallel hybridelectric vehicles, IEEE Transactions on Control Systems Technology12 (3) (2004) 352–363.[86] L. Serrao, S. Onori, G. Rizzoni, ECMS as a realization of pontryagin’sminimum principle for HEV control, Proceedings of the American Con-trol Conference (2009) 3964–3969.[87] N. Kim, S. Cha, H. Peng, Optimal control of hybrid electric vehiclesbased on Pontryagin’s minimum principle, IEEE Transactions on Con-trol Systems Technology 19 (5) (2011) 1279–1287.[88] C. Guardiola, B. Pla, S. Onori, G. Rizzoni, Insight into the HEV / PHEVoptimal control solution based on a new tuning method, Control Engi-neering Practice 29 (2014) 247–252.[89] A. Sciarretta, L. Serrao, P. C. Dewangan, P. Tona, E. N. Bergshoe ff ,C. Bordons, L. Charmpa, P. Elbert, L. Eriksson, T. Hofman,M. Hubacher, P. Isenegger, F. Lacandia, A. Laveau, H. Li, D. Marcos,T. N¨uesch, S. Onori, P. Pisu, J. Rios, E. Silvas, M. Sivertsson, L. Tribi-oli, A. J. van der Hoeven, M. Wu, A control benchmark on the energy anagement of a plug-in hybrid electric vehicle, Control EngineeringPractice 29 (2014) 287–298.[90] C. Musardo, G. Rizzoni, Y. Guezennec, B. Staccia, A-ECMS: AnAdaptive Algorithm for Hybrid Electric Vehicle Energy Management,European Journal of Control 11 (4-5) (2005) 509–524.URL http://linkinghub . elsevier . com/retrieve/pii/S0947358005710487 [91] D. Ambuhl, L. Guzzella, Predictive reference signal generator for hybridelectric vehicles, IEEE Transactions on Vehicular Technology 58 (9)(2009) 4730–4740.[92] P. Pisu, E. Silani, G. Rizzoni, S. M. Savaresi, A LMI-based SupervisoryRobust Control for Hybrid Vehicles, 2003 American Control Conference(2003) 4681–4686.[93] C.-C. Lin, H. Peng, J. W. Grizzle, A stochastic control strategy forhybrid electric vehicles, Proceedings of the 2004 American ControlConference 5 (2004) 4710–4715.URL http://ieeexplore . ieee . org/abstract/document/1384056/ [94] E. T. J. Dean, J. W. Grizzle, H. Peng, Shortest path stochastic control forhybrid electric vehicles, International Journal of Robust and NonlinearControl 18 (December 2007) (2007) 1409–1429.[95] D. F. Opila, X. Wang, R. McGee, R. B. Gillespie, J. A. Cook, J. W. Griz-zle, An energy management controller to optimally trade o ff fuel econ-omy and drivability for hybrid vehicles, IEEE Transactions on ControlSystems Technology 20 (6) (2012) 1490–1505.[96] A. A. Malikopoulos, A Multiobjective Optimization Framework forOnline Stochastic Optimal Control in Hybrid Electric Vehicles, IEEETransactions on Control Systems Technology 24 (2) (2015) 1–1.URL http://ieeexplore . ieee . org/document/7174519/ [97] M. L. Shaltout, A. A. Malikopoulos, S. Pannala, D. Chen, A Consumer-Oriented Control Framework for Performance Analysis in Hybrid Elec-tric Vehicles, IEEE Transactions on Control Systems Technology 23 (4)(2015) 1451–1464.[98] H. A. Borhan, C. Zhang, A. Vahidi, A. M. Phillips, M. L. Kuang,S. Di Cairano, Nonlinear model predictive control for power-split hy-brid electric vehicles, IEEE Conference on Decision and Control (2010)4890–4895.[99] H. Borhan, S. Member, A. Vahidi, A. M. Phillips, M. L. Kuang, I. V. Kol-manovsky, S. D. Cairano, MPC-Based Energy Management of a Power-Split Hybrid Electric Vehicle, IEEE Transactions on Control SystemsTechnology 20 (3) (2012) 593–603.[100] C. Manzie, T. S. Kim, R. Sharma, Optimal use of telemetry by paral-lel hybrid vehicles in urban driving, Transportation Research Part C 25(2012) 134–151.[101] C. Sun, S. J. Moura, X. Hu, J. K. Hedrick, F. Sun, Dynamic Tra ffi cFeedback Data Enabled Energy Management in Plug-in Hybrid Elec-tric Vehicles, IEEE Transactions on Control Systems Technology 23 (3)(2015) 1075–1086.[102] G. Ripaccioli, D. Bernardini, S. D. Cairano, A. Bemporad, I. Kol-manovsky, A stochastic model predictive control approach for series hy-brid electric vehicle power management, American Control Conference(ACC), 2010 (2010) 5844–5849.[103] M. Bichi, G. Ripaccioli, S. Di Cairano, D. Bernardini, A. Bemporad,I. Kolmanovsky, Stochastic model predictive control with driver behav-ior learning for improved powertrain control, 49th IEEE Conference onDecision and Control (2010) 6077–6082.[104] S. Di Cairano, D. Bernardini, A. Bemporad, I. V. Kolmanovsky, Stochas-tic MPC with learning for driver-predictive vehicle control and its ap-plication to HEV energy management, IEEE Transactions on ControlSystems Technology 22 (3) (2014) 1018–1031.[105] H. Wang, I. Kolmanovsky, M. R. Amini, J. Sun, Model Predictive Cli-mate Control of a Passenger Car in a Connected and Automated VehiclesEnvironment (2018). arXiv:1803 . .[106] S. Di Cairano, W. Liang, I. Kolmanovsky, M. Kuang, A. Phillips, En-gine Power Smoothing Energy Management Strategy for a Series HybridElectric Vehicle, 2011 American Control Conference - ACC 2011 (2)(2011) 2101–2106.[107] P. Pisu, K. Koprubasi, G. Rizzoni, Energy Management and DrivabilityControl Problems for Hybrid Electric Vehicles, in: 44th IEEE Confer-ence on Decision and Control, 2005, pp. 1824–1830.[108] T. N¨uesch, P. Elbert, M. Flankl, C. Onder, L. Guzzella, Convex optimiza- tion for the energy management of hybrid electric vehicles consideringengine start and gearshift costs, Energies 7 (2) (2014) 834–856.[109] L. Serrao, S. Onori, A. Sciarretta, Y. Guezennec, G. Rizzoni, Optimalenergy management of hybrid electric vehicles including battery aging,American Control Conference (ACC), 2011 (3) (2011) 2125–2130.[110] S. Ebbesen, P. Elbert, L. Guzzella, Battery State-of-Health PerceptiveEnergy Management for Hybrid Electric Vehicles, IEEE Trans. Veh.Technol. 61 (7) (2012) 2893–2900.[111] S. J. Moura, J. L. Stein, H. K. Fathy, Battery-Health Conscious PowerManagement in Plug-In Hybrid Electric Vehicles via ElectrochemicalModeling and Stochastic Control, Control Systems Technology, IEEETransactions on 21 (3) (2013) 679–694.[112] J. Guanetti, S. Formentin, S. Savaresi, Total cost minimization for nextgeneration Hybrid Electric Vehicles, in: IFAC Proceedings Volumes(IFAC-PapersOnline), Vol. 19, 2014.[113] S. Formentin, J. Guanetti, S. M. Savaresi, Least costly energy manage-ment for series hybrid electric vehicles, Control Engineering Practice 48(2016) 37–51.[114] J. Guanetti, S. Formentin, S. M. Savaresi, Least costly energy manage-ment for Electric Vehicles with plug-in Range Extenders, in: 2015 IEEE54th Conference on Decision and Control, 2015, pp. 638–643.[115] J. Guanetti, S. Formentin, S. Savaresi, Energy Management Systemfor an Electric Vehicle with a Rental Range Extender: A Least CostlyApproach, IEEE Transactions on Intelligent Transportation SystemsPP (99).[116] K. Yu, J. Yang, D. Yamaguchi, Model predictive control for hybrid ve-hicle ecological driving using tra ffi c signal and road slope information,Control Theory and Technology 13 (1) (2015) 17–28.[117] D. P. Bertsekas, Dynamic programming and optimal control, AthenaScientific, Belmont, MA, 1995.[118] B. Asadi, A. Vahidi, Predictive cruise control: Utilizing upcoming tra ffi csignal information for improving fuel economy and reducing trip time,IEEE Transactions on Control Systems Technology 19 (3) (2011) 707–714.[119] F. Lattemann, K. Neiss, S. Terwen, T. Connolly, The Predictive CruiseControl A System to Reduce Fuel Consumption of Heavy Duty Trucks,SAE Technical Paper 2004-01-26 (724).[120] H. Winner, S. Witte, W. Uhler, B. Lichtenberg, Adaptive Cruise ControlSystem Aspects and Development Trends (961010) (1996) 12.[121] L. Xiao, F. Gao, A comprehensive review of the development of adaptivecruise control systems, Vehicle System Dynamics 48 (10) (2010) 1167–1192.[122] D. Q. Mayne, J. B. Rawlings, C. V. Rao, P. O. Scokaert, Constrainedmodel predictive control: Stability and optimality, Automatica 36 (6)(2000) 789–814.[123] S. Lefevre, A. Carvalho, F. Borrelli, A Learning-Based Framework forVelocity Control in Autonomous Driving, IEEE Transactions on Au-tomation Science and Engineering (November) (2015) 1558–3783.[124] A. Vahidi, A. Eskandarian, Research advances in intelligent collisionavoidance and adaptive cruise control, IEEE Transactions on IntelligentTransportation Systems 4 (3) (2003) 143–153.[125] A. Higashimata, K. Adachi, T. Hashizume, S. Tange, Design of a head-way distance control system for ACC, JSAE review 22 (1) (2001) 15–22.[126] V. Turri, Y. Kim, J. Guanetti, K. H. Johansson, F. Borrelli, A model pre-dictive controller for non-cooperative eco-platooning, in: 2017 Ameri-can Control Conference, Seattle, 2017, pp. 2309–2314.[127] A. Al Alam, J. Mårtensson, K. H. Johansson, Experimental evaluationof decentralized cooperative cruise control for heavy-duty vehicle pla-tooning, Control Engineering Practice 38 (2015) 11–25.[128] M. Zabat, N. Stabile, S. Farascaroli, F. Browand, The Aerodynamic Per-formance Of Platoons: A Final Report, California Partners for AdvancedTransit and Highways (PATH).[129] M. Barth, S. Mandava, K. Boriboonsomsin, H. Xia, Dynamic ECO-driving for arterial corridors, 2011 IEEE Forum on Integrated and Sus-tainable Transportation Systems, FISTS 2011 (2011) 182–188.[130] B. Van Arem, C. J. Van Driel, R. Visser, The impact of cooperative adap-tive cruise control on tra ffi c-flow characteristics, IEEE Transactions onIntelligent Transportation Systems 7 (4) (2006) 429–436.[131] V. Milanes, S. E. Shladover, J. Spring, C. Nowakowski, H. Kawazoe,M. Nakamura, Cooperative adaptive cruise control in real tra ffi c situa-tions, IEEE Transactions on Intelligent Transportation Systems 15 (1) ffi cient and safe heavy-duty vehicle platooning 25 (1) (2015)12–28.[145] L. Bertoni, J. Guanetti, M. Basso, M. Masoero, S. Cetinkunt, F. Borrelli,An adaptive cruise control for connected energy-saving electric vehicles,in: IFAC World Congress, Vol. 50, 2017, pp. 2359–2364.[146] S. E. Li, K. Deng, Y. Zheng, H. Peng, E ff ect of Pulse-and-Glide Strategyon Tra ffi c Flow for a Platoon of Mixed Automated and Manually DrivenVehicles, Computer-Aided Civil and Infrastructure Engineering 30 (11)(2015) 892–905.[147] K. Mcdonough, I. Kolmanovsky, D. Filev, D. Yanakiev, S. Szwabowski,J. Michelini, Stochastic Dynamic Programming Control Policies forFuel E ffi cient Vehicle Following, in: American Control Conference,2013, pp. 1350–1355.[148] R. Schmied, H. Waschl, L. Del Re, Extension and experimental vali-dation of fuel e ffi cient predictive adaptive cruise control, in: AmericanControl Conference, Vol. 2015-July, 2015, pp. 4753–4758.[149] D. Moser, H. Waschl, H. Kirchsteiger, R. Schmied, L. del Re, Coop-erative adaptive cruise control applying stochastic linear model predic-tive control strategies, in: 2015 European Control Conference, 2015, pp.3383–3388.[150] R. Schmied, H. Waschl, R. Quirynen, M. Diehl, L. Del Re, NonlinearMPC for Emission E ffi cient Cooperative Adaptive Cruise Control, in:IFAC-PapersOnLine, Vol. 48, Elsevier B.V., 2015, pp. 160–165.[151] G. J. Naus, R. P. Vugts, J. Ploeg, M. J. Van De Molengraft, M. Steinbuch,String-stable CACC design and experimental validation: A frequency-domain approach, IEEE Transactions on Vehicular Technology 59 (9)(2010) 4268–4279.[152] J. L. G. Naus, R. P. A. Vugts, J. Ploeg, M. J. G. van de Molengraft,M. Steinbuch, String-Stable CACC Design and Experimental Validation,IEEE Transactions of Vehicular Technology 59 (9) (2010) 4268–4279. [153] W. B. Dunbar, D. S. Caveney, Distributed receding horizon control ofvehicle platoons: Stability and string stability, IEEE Transactions onAutomatic Control 57 (3) (2012) 620–633.[154] M. Diaby, A. Sorkati, Optimization for Energy E ffi cient CooperativeAdaptive Cruise Control, Ph.D. thesis (2016).[155] Y. Zheng, S. E. Li, K. Li, F. Borrelli, J. K. Hedrick, Distributed ModelPredictive Control for Heterogeneous Vehicle Platoons under Unidirec-tional Topologies, IEEE Transactions on Control Systems Technology25 (3) (2017) 899–910.[156] F. Borrelli, P. Falcone, T. Keviczky, J. Asgari, D. Hrovat, MPC-based ap-proach to active steering for autonomous vehicle systems, InternationalJournal of Vehicle Autonomous Systems 3 (2 / /
4) (2005) 265.[157] P. Falcone, M. Tufo, F. Borrelli, J. Asgari, H. E. Tseng, A linear timevarying model predictive control approach to the integrated vehicle dy-namics control problem in autonomous systems, 2007 46th IEEE Con-ference on Decision and Control (2007) 2980–2985.[158] V. Turri, A. Carvalho, H. E. Tseng, K. H. Johansson, F. Borrelli, Lin-ear model predictive control for lane keeping and obstacle avoidanceon low curvature roads, IEEE Conference on Intelligent TransportationSystems, Proceedings, ITSC (2013) 378–383.[159] A. Carvalho, S. Lef´evre, G. Schildbach, J. Kong, F. Borrelli, Automateddriving: The role of forecasts and uncertainty - A control perspective,European Journal of Control 24 (2015) 14–32.[160] R. White, M. Tomizuka, Autonomous following lateral control of heavyvehicles using laser scanning radar, American Control Conference,2001. Proceedings of the 2001 3 (2001) 2333–2338.[161] S. K. Gehrig, F. J. Stein, A Trajectory-Based Approach for the LateralControl of Car Following Systems, in: IEEE International Conferenceon Systems, Man, and Cybernetics, 1998, pp. 3596–3601.[162] Z. Huang, Q. Wu, J. Ma, S. Fan, An APF and MPC combined collab-orative driving controller using vehicular communication technologies,Chaos, Solitons and Fractals 89 (2015) 232–242.[163] S. Lefevre, A. Carvalho, F. Borrelli, Autonomous car following: Alearning-based approach, IEEE Intelligent Vehicles Symposium, Pro-ceedings 2015-Augus (Iv) (2015) 920–926.[164] C. Katrakazas, M. Quddus, W.-H. Chen, L. Deka, Real-time motionplanning methods for autonomous on-road driving: State-of-the-art andfuture research directions, Transportation Research Part C: EmergingTechnologies 60 (2015) 416–442.[165] C. Laugier, J. Iba, C. Laugier, J. Iba, S. Lef`evre, C. Laugier, J. Iba˜nez-guzm´an, Risk Assessment at Road Intersections : Comparing Intentionand Expectation St To cite this version : Risk Assessment at Road In-tersections : Comparing Intention and Expectation, Intelligent VehiclesSymposium (IV), 2012 IEEE (2012) 165–171.[166] X. Zhang, A. Liniger, F. Borrelli, Optimization-Based Collision Avoid-ance (2017) 1–24.URL http://arxiv . org/abs/1711 . [167] J. Kong, M. Pfei ff er, G. Schildbach, F. Borrelli, Kinematic and DynamicVehicle Models for Autonomous Driving Control Design, IEEE Intelli-gent Vehicles Symposium (Iv) (2015) 1094–1099.[168] A. Dardanelli, M. Tanelli, B. Picasso, S. M. Savaresi, O. Di Tanna,M. D. Santucci, A smartphone-in-the-loop active state-of-charge man-ager for electric vehicles, IEEE / ASME Transactions on Mechatronics17 (3) (2012) 454–463.[169] P. Sharer, A. Rousseau, D. Karbowski, S. Pagerit, Plug-in Hybrid Elec-tric Vehicle Control Strategy: Comparison between EV and Charge-Depleting Options, in: SAE, 2008.[170] P. Tulpule, V. Marano, G. Rizzoni, E ff ects of di ff erent PHEV controlstrategies on vehicle performance, in: 2009 American Control Confer-ence, 2009, pp. 3950–3955.[171] J. Guanetti, S. Formentin, M. Corno, S. M. Savaresi, Optimal energymanagement in series hybrid electric bicycles, Automatica 81 (2017)96–106.[172] V. Larsson, Route Optimized Energy Management of Plug-in HybridElectric Vehicles, Ph.D. thesis, Chalmers University of Technology(2014).[173] V. Larsson, L. Johannesson Mardh, B. Egardt, L. J. Mårdh, B. Egardt,Comparing two approaches to precompute discharge strategies for plug-in hybrid electric vehicles, in: 7th IFAC Symposium on Advances inAutomotive Control, Vol. 7, IFAC, 2013, pp. 121–126.[174] C. Manzie, P. Dewangan, G. Corde, O. Grondin, A. Sciarretta, State of harge Management for Plug-In Hybrid Vehicles With Uncertain TripInformation, Journal of Dynamic Systems, Measurement, and Control137 (9) (2015) 091005.[175] D. E. Kirk, Optimal Control Theory, Dover Publications, Mineola, NewYork, 1998.[176] E. Ozatay, S. Onori, J. Wollaeger, U. Ozguner, G. Rizzoni, D. Filev,J. Michelini, S. Di Cairano, Cloud-based velocity profile optimizationfor everyday driving: A dynamic-programming-based solution, IEEETransactions on Intelligent Transportation Systems 15 (6) (2014) 2491–2505.[177] B. Asadi, A. Vahidi, Predictive use of tra ffi c signal state for fuel saving,Vol. 42, IFAC, 2009.[178] K. Katsaros, Performance study of a Green Light Optimal Speed Advi-sory ( GLOSA ) Application Using an Integrated Cooperative ITS Sim-ulation Platform, Proceedings of Wireless Communications and MobileComputing Conference (IWCMC) (2011) 918–923.[179] E. Koukoumidis, M. Martonosi, L. S. Peh, Leveraging smartphone cam-eras for collaborative road advisories, IEEE Transactions on MobileComputing 11 (5) (2012) 707–723.[180] G. Mahler, A. Vahidi, Reducing idling at red lights based on probabilisticprediction of tra ffi c signal timings, American Control Conference (ACC)(2012) 6557 – 6562.[181] S. Mandava, K. Boriboonsomsin, M. Barth, Arterial velocity planningbased on tra ffi c signal information under light tra ffi c conditions, IEEEConference on Intelligent Transportation Systems, Proceedings, ITSC(2009) 160–165.[182] M. Miyatake, M. Kuriyama, Y. Takeda, Theoretical study on eco-drivingtechnique for an electric vehicle considering tra ffi c signals, in: Interna-tional Conference on Power Electronics and Drive Systems, no. Decem-ber, 2011, pp. 733–738.[183] G. De Nunzio, C. Canudas de Wit, P. Moulin, D. Di Domenico, Eco-driving in urban tra ffi c networks using tra ffi c signal information, 52ndIEEE Conference on Decision and Control (2013) 892–898.[184] G. De Nunzio, C. C. d. Wit, P. Moulin, D. D. Domenico, Eco-drivingin urban tra ffi c networks using tra ffi c signals information, InternationalJournal of Robust and Nonlinear Control 26 (2016) 1307–1324.[185] B. HomChaudhuri, A. Vahidi, P. Pisu, Fast Model Predictive Control-Based Fuel E ffi cient Control Strategy for a Group of Connected Vehi-cles in Urban Road Conditions, IEEE Transactions on Control SystemsTechnology 25 (2) (2017) 760–767.[186] B. HomChaudhuri, A. Vahidi, P. Pisu, A fuel economic model predic-tive control strategy for a group of connected vehicles in urban roads,American Control Conference (2015) 2741–2746.[187] M. A. S. Kamal, M. Mukai, J. Murata, T. Kawabe, On Board Eco -Driving System for Varying Road - Tra ffi c Environments Using ModelPredictive Control (2010) 1636–1641.[188] D. YAMAGUCHI, M. KAMAL, M. MUKAI, T. KAWABE, Model Pre-dictive Control for Automobile Ecological Driving Using Tra ffi c SignalInformation, Journal of System Design and Dynamics 6 (3) (2012) 297–309.[189] M. Seredynski, W. Mazurczyk, D. Khadraoui, Multi-segment green lightoptimal speed advisory, Proceedings - IEEE 27th International Paral-lel and Distributed Processing Symposium Workshops and PhD Forum,IPDPSW 2013 (2013) 459–465.[190] E. Ozatay, U. Ozguner, D. Filev, J. Michelini, Analytical and numericalsolutions for energy minimization of road vehicles with the existenceof multiple tra ffi c lights, in: 52nd IEEE Conference on Decision andControl, 2013, pp. 7137–7142.[191] H. Xia, K. Boriboonsomsin, F. Schweizer, A. Winckler, K. Zhou, W. B.Zhang, M. Barth, Field operational testing of ECO-approach technologyat a fixed-time signalized intersection, IEEE Conference on IntelligentTransportation Systems, Proceedings, ITSC (2012) 188–193.[192] M. Li, K. Boriboonsomsin, G. Wu, W.-B. Zhang, M. Barth, Tra ffi c En-ergy and Emission Reductions at Signalized Intersections : A Study ofthe Benefits of Advanced Driver Information, International Journal ofITS Research 7 (1) (2009) 49–58.[193] H. Xia, G. Wu, K. Boriboonsomsin, M. J. Barth, Development and Eval-uation of an Enhanced Eco - Approach Tra ffi c Signal Application forConnected Vehicles, 16th International IEEE Conference on IntelligentTransportation Systems (ITSC 2013) (Itsc) (2013) 296–301.[194] C. Sun, J. Guanetti, F. Borrelli, S. Moura, Robust Eco-Driving Control of Autonomous Vehicles Connected to Tra ffi c Lights (2018) 1–14.URL http://arxiv . org/abs/1802 . [195] J. Hellendoorn, B. De Schutter, L. Baskar, Z. Papp, Tra ffi c control andintelligent vehicle highway systems: a survey, IET Intelligent TransportSystems 5 (1) (2011) 38–52.[196] L. Li, D. Wen, D. Yao, A survey of tra ffi c control with vehicular com-munications, IEEE Transactions on Intelligent Transportation Systems15 (1) (2014) 425–432.[197] B. Khondaker, L. Kattan, Variable speed limit: A microscopic analysisin a connected vehicle environment, Transportation Research Part C:Emerging Technologies 58 (2015) 146–159.[198] A. A. Malikopoulos, S. Hong, J. Lee, B. B. Park, Development and Eval-uation of Speed Harmonization Using Optimal Control Theory (2016).URL http://arxiv . org/abs/1611 . [199] A. A. Malikopoulos, S. Hong, J. Lee, B. B. Park, Optimal Speed Controlof Automated Vehicles at Speed Reduction Zones (2016).URL http://arxiv . org/abs/1611 . [200] J. Rios-Torres, A. A. Malikopoulos, A Survey on the Coordination ofConnected and Automated Vehicles at Intersections and Merging atHighway On-Ramps, IEEE Transactions on Intelligent TransportationSystems 18 (5) (2017) 1066–1077.[201] J. Rios-Torres, A. A. Malikopoulos, Automated and Cooperative Vehi-cle Merging at Highway On-Ramps, IEEE Transactions on IntelligentTransportation Systems 18 (4) (2017) 780–789.[202] G. Guo, W. Yue, Hierarchical platoon control with heterogeneous in-formation feedback, IET Control Theory & Applications 5 (15) (2011)1766–1781.[203] S. Solyom, E. Coelingh, Performance limitations in vehicle platoon con-trol, IEEE Intelligent Transportation Systems Magazine 5 (4) (2013)112–120.[204] R. Hall, C. Chin, Vehicle sorting for platoon formation: Impacts on high-way entry and throughput, Transportation Research Part C: EmergingTechnologies 13 (5-6) (2005) 405–420.[205] J. Frankel, L. Alvarez, R. Horowitz, P. Li, Safety Oriented Maneuversfor IVHS, Vehicle System Dynamics 26 (4) (2007) 271–299.[206] M. Goli, A. Eskandarian, A systematic multi-vehicle platooning and pla-toon merging: Strategy, control, and trajectory generation, in: ASMEDynamic Systems and Control Conference, 2014.[207] J. P. Koller, A. G. Colin, B. Besselink, K. H. Johansson, Fuel-E ffi cientControl of Merging Maneuvers for Heavy-Duty Vehicle Platooning,IEEE Conference on Intelligent Transportation Systems, Proceedings,ITSC 2015-Octob (2015) 1702–1707.[208] K.-Y. Liang, Coordination and Routing for Fuel-E ffi cient Heavy-DutyVehicle Platoon Formation, 2014.[209] B. Besselink, V. Turri, S. H. Van De Hoef, K. Y. Liang, A. Al Alam,J. M ˜Artensson, K. H. Johansson, Cyber-Physical Control of RoadFreight Transport, Proceedings of the IEEE 104 (5) (2016) 1128–1141.[210] Q. Jin, G. Wu, K. Boriboonsomsin, M. Barth, S. Member, Platoon -Based Multi - Agent Intersection Management for Connected Vehicle,16th International IEEE Conference on Intelligent Transportation Sys-tems (ITSC 2013) (Itsc) (2013) 1462–1467.[211] A. Lelouvier, J. Guanetti, F. Borrelli, Eco-Platooning of AutonomousElectrical Vehicles Using Distributed Model Predictive Control, in:IEEE 20th Intelligent Transportation Systems Conference, 2017, pp.464–469.[212] M. Barth, K. Boriboonsomsin, A. Vu, Environmentally-Friendly naviga-tion, IEEE Conference on Intelligent Transportation Systems, Proceed-ings, ITSC (2007) 684–689.[213] M. Barth, C. Malcolm, T. Younglove, N. Hill, Recent validation e ff ortsfor a comprehensive modal emissions model, Transportation ResearchRecord: Journal of the Transportation Research Board 1750 (-1) (2001)13–23.[214] K. Boriboonsomsin, M. J. Barth, W. Zhu, A. Vu, Eco-routing navigationsystem based on multisource historical and real-time tra ffi c information,IEEE Transactions on Intelligent Transportation Systems 13 (4) (2012)1694–1704.[215] O. Andersen, C. S. Jensen, K. Torp, B. Yang, EcoTour: Reducing the en-vironmental footprint of vehicles using eco-routes, Proceedings - IEEEInternational Conference on Mobile Data Management 1 (2013) 338–340.[216] T. Jurik, A. Cela, R. Hamouche, R. Natowicz, A. Reama, S. I. Niculescu, . Julien, Energy optimal real-time navigation system, IEEE IntelligentTransportation Systems Magazine 6 (3) (2014) 66–79.[217] E. Yao, Y. Song, Study on eco-route planning algorithm and environ-mental impact assessment, Journal of Intelligent Transportation Sys-tems: Technology, Planning, and Operations 17 (1) (2013) 42–53.[218] B. Yang, C. Guo, C. S. Jensen, M. Kaul, S. Shang, Stochastic skylineroute planning under time-varying uncertainty, Proceedings - Interna-tional Conference on Data Engineering (2014) 136–147.[219] C. Guo, B. Yang, O. Andersen, C. S. Jensen, K. Torp, EcoSky: Reduc-ing vehicular environmental impact through eco-routing, Proceedings -International Conference on Data Engineering 2015-May (2015) 1412–1415.[220] M. Kubicka, J. Klusacek, A. Sciarretta, A. Cela, H. Mounier,L. Thibault, S. I. Niculescu, Performance of current eco-routingmethods, IEEE Intelligent Vehicles Symposium, Proceedings 2016-Augus (Iv) (2016) 472–477.[221] D. Krajzewicz, G. Hertkorn, SUMO (Simulation of Urban MObility) Anopen-source tra ffi c simulation, . . . Symposium on Simulation . . . (2002)63–68.[222] Z. Sun, X. Zhou, To save money or to save time: Intelligent routingdesign for plug-in hybrid electric vehicle, Transportation Research PartD: Transport and Environment 43 (2011) (2016) 238–250.[223] S. Pourazarm, C. G. Cassandras, Optimal Routing of Energy-awareVehicles in Networks with Inhomogeneous Charging Nodes, in: 22ndMediterranean Conference on Control and Automation, 2014, pp. 674–679.[224] T. Wang, C. G. Cassandras, S. Pourazarm, Energy-aware Vehicle Rout-ing in Networks with Charging Nodes, in: 19th IFAC World Congress,IFAC, 2014, pp. 9611–9616.[225] S. Pourazarm, C. G. Cassandras, A. Malikopoulos, Optimal routing ofelectric vehicles in networks with charging nodes: A dynamic program-ming approach, in: 2014 IEEE International Electric Vehicle Confer-ence, 2015.[226] J. Sun, H. X. Liu, Stochastic Eco-routing in a Signalized Tra ffi c Net-work, Transportation Research Procedia 7 (2015) 110–128.c Net-work, Transportation Research Procedia 7 (2015) 110–128.