A Fleet of Miniature Cars for Experiments in Cooperative Driving
aa r X i v : . [ c s . R O ] F e b A Fleet of Miniature Cars for Experiments in Cooperative Driving
Nicholas Hyldmar ∗ , Yijun He ∗ , Amanda Prorok Abstract — We introduce a unique experimental testbed thatconsists of a fleet of 16 miniature Ackermann-steering vehicles.We are motivated by a lack of available low-cost platformsto support research and education in multi-car navigationand trajectory planning. This article elaborates the designof our miniature robotic car, the
Cambridge Minicar , as wellas the fleet’s control architecture. Our experimental testbedallows us to implement state-of-the-art driver models as wellas autonomous control strategies, and test their validity in areal, physical multi-lane setup. Through experiments on ourminiature highway, we are able to tangibly demonstrate thebenefits of cooperative driving on multi-lane road topographies.Our setup paves the way for indoor large-fleet experimentalresearch.
I. I
NTRODUCTION
The deployment of connected, automated, and autonomousvehicles presents us with transformational opportunities forroad transport. To date, the number of companies work-ing on this technology is substantive, and growing [1].Opportunities reach beyond single-vehicle automation: byenabling groups of vehicles to jointly agree on maneuversand navigation strategies, real-time coordination promises toimprove overall traffic throughput, road capacity, and passen-ger safety [5, 6]. However, coordinated driving for intelligentvehicles still remains a challenging research problem, due tounpredictable vehicle behaviors (e.g., non-cooperative cars,unreliable communication), hard workspace limitations (e.g.,lane topographies), and constrained kinodynamic capabilities(e.g., steering kinematics, driver comfort).Developing true-scale facilities for safe, controlled vehicletestbeds is massively expensive and requires a vast amount ofspace. For example, the University of Michigan’s MCity TestFacility cost US $10 million to develop and covers 32 acres(0.13 km ). As a consequence, using fleets of actual vehiclesis possible only for very few research institutes worldwide.Moreover, although such facilities are excellently suited forresearch and development, safety and operational concernsprohibit the integration of educational curricula and outreachactivities. One approach to facilitating experimental researchand education is to build low-cost testbeds that incorporatefleets of down-sized, car-like mobile platforms. Followingthis idea, we propose a multi-car testbed that allows for theoperation of tens of vehicles within the space of a moderatelylarge robotics laboratory, and allows for the teaching andresearch of coordinated driving strategies in dense trafficscenarios.Our motivation is to design a testbed that scales to alarge number of cars so that we could test vehicle-to-vehicle All authors are with the University of Cambridge, UK: { nh490,yh403, asp45 } @cam.ac.uk . ∗ These authors contributed equallyto this work. We gratefully acknowledge the Isaac Newton Trust who aresupporting Amanda Prorok through an Early Career Grant.
Fig. 1:
The fleet of Minicars on a U-shaped two-lane miniature freeway.The inner and outer track lengths are 16 m and 17 m, respectively. interactions (cooperative as well as non-cooperative) andthe effect of these interactions in multi-car traffic scenarios.Although a number of low-cost multi-robot testbeds exist(e.g., [8, 17]), most use robotic platforms with differentialdrive kinematics (which tend to also have limited maximumspeeds). Few testbeds integrate car-like Ackermann-steeringvehicles, and in very small numbers (e.g., [12]).In this work, we propose the design of a low-cost miniaturerobotic car, the
Cambridge Minicar , which is based on a 1:24model of an existing commercial car. The Minicar is builtfrom off-the-shelf components (with the exception of onelaser-cut piece), and costs approximately US $76 in its basicconfiguration. Its low cost allows us to compose a large fleet,which we use to test navigation strategies and driver models.Overall, our contributions in this work are (a) the design ofa low-cost miniature robotic car, (b) the proposition of asystem architecture that incorporates decentralized multi-carcontrol algorithms for indoor testing of real large fleets, and (c) the availability of our designs and code in an open-sourcerepository . II. E XISTING P LATFORMS
A variety of low-cost mobile robot platforms are availablefor research and education. The work in [16] presents avery recent comprehensive overview of platforms that costless than US $ 300, have been designed in the last 10years, or are currently available on the market. Apart fromthe Kilobot [20] and the AERobot [21], which have slip-stick forwards motion, all other robots in this overview arewheeled differential drive platforms [2–4, 10, 15, 16, 19, 24].The lack of low-cost, Ackermann-steering robots is apparent.There are a few recent robot designs that are based onAckermann-steering platforms, and we provide an overviewthereof in Table I. The largest of these platforms is the https://github.com/proroklab/minicar ar Scale Price (USD) ETHZ ORCA Racer [14] 1:43 $ 470
Cambridge Minicar
TABLE I:
Overview of Ackermann-steering platforms. Georgia Tech AutoRally [23] . It is based on a 1:5-scalevehicle chassis, runs on an ASUS Mini-ITX motherboardthat includes a GPU, and is capable of fast autonomous driv-ing using on-board sensors only. It is specifically designedfor outdoor experimentation and the testing of aggressivemaneuvers (all computational components are housed ina rugged aluminum enclosure able to withstand violentvehicle rollovers.) The MIT Racecar [9] and the BerkeleyAutonomous Race Car (BARC) [7] are 1:10-scale rally carsbased on a chassis that is commercially available (Traxxas).Similar to the AutoRally, these platforms feature high-endcomputational units with GPUs that allow for the addition ofnumerous sensors (e.g., laser range finder, camera, GPS) foron-board autonomy. As listed in Table I, the price of thesethree platforms lies in the range $ 800-$ 9300 for the corealone (chassis, motors, and computational units). Due to thesignificant cost and moderately large size of the platforms, itis difficult to facilitate indoor experiments that include largenumbers of these vehicles.At the other end of the spectrum lies the ETHZ ORCARacer [14] , which is based on a 1:43-scale platform. Thechassis is provided by a Kyosho dnano RC race car (achiev-ing top speeds of more than 3 m/s). In order to control theplatform, the authors designed a custom PCB with an ARMCortex M4 microcontroller, a Bluetooth chip, and H-bridgesfor DC motor actuation (for the steering servo and drivetrain). The overall cost of one car is $ 470 (approximately$ 300 for the PCB parts and $ 170 for the dnano chassis).This platform provides interesting dynamic capabilities, anda large number of cars can be operated in very confinedspaces. Although the platform is considerably cheaper thanthe aforementioned models, it is still significantly moreexpensive than our Minicar (with similar capabilities). Fur-thermore, the platform is too small to carry an off-the-shelfcomputer, such as Raspberry Pi Zero W (which facilitatescommunication via WiFi, as well as the addition of varioussensors). Finally, its design and accompanying software toolsare not open-sourced. These factors limit the extensibility andscalability of the testbed.III. V EHICLE D ESIGN
The main components of the Minicar are a RaspberryPi Zero W, a chassis with forwards drive train and servo-motors, and two battery sets. Figure 2 shows an exploded https://autorally.github.io/ https://github.com/mit-racecar https://sites.google.com/site/orcaracer/home This overview lists platforms developed in the last 10 years. Prices in-dicate the cost of the core body, including chassis, motors and computationalunits (excluding additional sensor components.)
Component Item Price (USD)
Computation Raspberry Pi Zero W $ 12.3Memory 8GB Micro SD card $ 5.2Chassis & Motor 1:24 Range Rover Sport $ 15.6Steering Micro-servo $ 6.2Motor Power AA Batteries $ 11.1Logic Power Portable Charger $ 15.6Boost converter XL6009 $ 2.6H-Bridge L293D $ 0.5Board Proto Bonnet $ 4.5Logic Switch USB Switch $ 2.9
TABLE II:
Cambridge Minicar: overview of components.
Fig. 2:
Exploded view of a Cambridge Minicar. 1: Headlights 2: Portablecharger 3: Casing screws 4: Gear 5: Servo 6: Lower casing 7: H-bridge8: Micro SD card 9: USB switch 10: Circuit board 11: Upper casing 12: JSTconnectors 13: Micro USB cable 14: Boost converter 15: Raspberry Pi ZeroW 16: Motor switch 17: Drive motor view of the Minicar. It is × × mm and weighs450 g (including batteries). The logic can be powered forover 5 hours and the motors for 2.2 hours at 0.3 m/s. Thelogic is powered separately from the motors to isolate itfrom the noisy environment and increase the car’s runtime.The motor and logic powers have capacities of 2500 mAhand 3350 mAh respectively. The AA batteries supply 4.2 Vto the servo which is increased to 7 V for the motor bythe boost converter. The servo logic and motor enable pinare controlled using pulse width modulation (PWM) by thePi Zero W. A servo arm is connected to a laser-cut gearthat meshes with the existing steering gear. The vehicle’swheel base is L = 122 mm. Its minimum turning radiusis R = | ψ | max =18 ◦ , its maximum steering rate is | ˙ ψ | max = 0 . rad/sand its maximum forwards speed is v max = ESTBED
The architecture of our system is illustrated in Figure 3.We use an OptiTrack motion capture system based on tate Estimation Motion CaptureSystemTrajectoryPlanner Controller n e i ghbo r i ng ca r s setpoint motor commands (100Hz) (50 Hz) pose array e go ca r Radio
Fig. 3:
Diagram of our system architecture. passive reflective markers to provide real-time feedback onthe vehicles’ positions. Each Minicar is equipped with aunique configuration of five markers. The motion capturesystem tracks the Minicars and provides pose measurementsat 100 Hz. An extended Kalman filter uses the pose arrayto provide an estimate of each vehicle’s state that includesvehicle pose [ x, y, θ ] , velocity v and steering angle ψ . Weimplement an inner and outer loop control method fortrajectory tracking. The outer loop (i.e., Trajectory Planner)generates trajectories (or uses pre-computed trajectories, suchas freeway lanes) to compute velocity and steering anglesetpoints, which are fed to the inner loop. In Section VI,we provide an example of a Trajectory Planner for use inmulti-lane freeway traffic. The inner loop (i.e., Controller)is responsible for correcting motor commands using PIDcontrol on velocity and steering angle. These control valuesare sent to the vehicles over broadband radio. The on-boardcomputer applies the corresponding motor commands withpulse-with modulated signals. At the scale of our currentsystem (16 vehicles), we did not notice any communicationlatency. Should this become an issue at larger numbers ofvehicles, we note that the architecture can easily be scaledby considering multi-radio solutions, such as in [18].Our testbed architecture is designed for ease of use,and our key aim is the rapid development and testing ofdriving behaviors on car-like robots (such as the Minicar).Although the Minicar’s design allows for the integration ofsensors (such as an IMU or camera ), and its on-boardcomputer is capable of performing its own state estimationand control computations, we decided to keep the intelli-gence off-board, emulating proprioceptive and exteroceptiveobservations through the motion capture system instead. Inour setup, a workstation runs a thread for each Minicar’sTrajectory Planner and Controller. Upon booting, the Minicarexecutes a listener that waits for motor commands. Thissoftware design choice facilitates rapid testing, and allows usto focus on the development of driving strategies for largeMinicar fleets. Our fleet operates on a miniature two-laneU-shaped freeway, shown in Figure 1.V. T RAJECTORY T RACKING
The Minicar has Ackermann steering geometry, and itskinematics can be approximated by the bicycle model, withmotion equations as follows: ˙ x = v cos θ (1) ˙ y = v sin θ (2) ˙ θ = L − v tan ψ (3) A dedicated Raspberry Pi Camera Module costs less than $ 30; itsminiature form factor allows it to be easily fitted onto the Minicar.
PSfrag replacements x d QP y xθ ψ d ψ l l l x t Reference Trajectory R e f e r e n ce C a r R e a l C a r Fig. 4:
Illustration of our trajectory tracking strategy, based on [13]. where ψ is the steering angle, v is the forward speed, and L is the vehicle’s wheel base. We use the lateral controlstrategy introduced in [13]. This strategy has the advantageof being speed independent, so that the velocity of the vehiclecan be controlled independently for other purposes and stillconverge to the desired pose on the trajectory.Figure 4 illustrates the control strategy. The real vehiclewith yaw θ and position is projected orthogonally onto thereference trajectory to create a virtual reference vehicle atpoint x d , which is the closest point on the reference trajectoryto the real vehicle. The yaw of the virtual reference vehicleis aligned with the tangent at x d , and its steering angle ψ d is such that the vehicle’s turning radius coincides with thecurvature κ of the reference trajectory at x d . We computethe target point x t with x t = x d + l cos θ d + l cos( θ d + ψ d ) (4) y t = y d + l sin θ d + l sin( θ d + ψ d ) (5)where the steering angle of the reference vehicle is ψ d =arctan( l κ ) . The real vehicle’s steering angle is ψ = arctan2( y t − l sin θ, x t − l cos θ ) − θ. (6)As noted in [13], the choice of parameters l and l affectthe control; small values for l lead to fast control, but maylead to an overshoot of the reference trajectory; small valuesfor l (gain) lead to fast control, but also mean high controlvalues, which may lead to a saturation of the steering angle.We experimentally tuned our parameters, and set l = L and l = 2 . L .VI. A M ULTI -C AR T RAFFIC S YSTEM
Building on the system architecture and vehicle controlstrategy described above, we design an experimental multi-car multi-lane traffic system that emulates freeway drivingin an indoor lab setup. In our architecture, each vehicle iscontrolled by its individual, independent trajectory planningmodule, and hence, heterogeneous driving behaviors are pos-sible. Our aim is to show how our experimental setup allowsus to validate various driving controllers on actual platformsin a freeway-like setting. To this end, we implement threecontrol strategies: (i) the car is driven by an egocentric,human-like policy, (ii) the car is driven by a cooperativepolicy, and (iii) the car is driven by a human player (i.e., agamified policy). S / State EstimationLateral ControlCommunicationsdesiredsteering angle adjacent lanes same lane
C-MOBIL projected neighbors realneighbors lane change decision c o mm un i ca t e d e s i r e d l a n e desiredvelocityacceleration projected actual ego C-IDM
PSfrag replacements
Trajectory PlannerFig. 5:
Diagram of trajectory planner with our two main algorithmicmodules, C-MOBIL and C-IDM. The ego vehicle plans a trajectory andvelocity profile based on its own state, and on the actual state of itsneighboring vehicles. In a cooperative approach (with diagram elementsmarked in blue), the algorithm modules also integrate the projected (desired)states of neighboring vehicles.
A. Egocentric Driving
Our human-like driving model treats longitudinal andlateral control as two independent entities. It uses an ac-celeration model that controls longitudinal motion along thecurrent car lane, and a steering model that controls lateralmotion across multiple lanes.
Longitudinal Control.
Our longitudinal control is basedon the Intelligent Driver Model (IDM) first proposed in [22].The idea underpinning IDM is that a vehicle’s accelerationis a function of the vehicle’s current velocity v , it’s gap s to the preceding vehicle, and the approach rate ∆ v to thepreceding vehicle. It is formalized as a IDM = α " − (cid:18) vv (cid:19) δ − (cid:18) s ⋆ ( v, ∆ v ) s (cid:19) (7)where s ⋆ is a function determining the desired minimum gapto the preceding vehicle. We compute this value as s ⋆ ( v, ∆ v ) = s + T v + v ∆ v √ αβ . (8)where s corresponds to a jam distance. The vehicle’s controlinput at time t is computed through the integrator v t = a IDM · ∆ t + v t − . Lateral Control.
Our lateral control strategy builds onthe MOBIL lane changing policy proposed in [11]. TheMOBIL strategy decides whether a vehicle should changea lane or not based on two steps. First, it ensures thata safety criterion is met, i.e., it guarantees that after thelane-change, the deceleration of the new follower vehicle a n does not exceed a safety limit, a n ≥ β n . Second, itvalidates via an incentive criterion that checks whether thelocal traffic situation of the car would improve, given a lanechange. This incentive model also involves the immediately IDM Parameter
Normal Aggressive
Desired velocity v [m/s] 0.4 0.4Time headway T [s] 2.0 2.0Max. acceleration α [m/s ] 0.5 1.0Desired deceleration β [m/s ] 0.3 0.5Acceleration exp. δ s [m] 0.1 0.1 MOBIL Parameter
Normal Aggressive
Politeness p β n [m/s ] . · α . · α Accel. thresh. ∆ a T [m/s ] 0.4 0.2 TABLE III:
Minicar parameter values for lateral and longitudinal control,in a normal and an aggressive mode. affected vehicles, i.e., the new and old follower vehicles.We evaluate the difference in acceleration for the currentvehicle ( ∆ a c ), the new follower ( ∆ a n ), and the old follower( ∆ a o ), whereby a positive acceleration change for the currentcar is considered favorable. A politeness factor p ∈ [0 , determines how much heed is payed to the new and oldfollower vehicles. The incentive to change lanes is controlledby a switching threshold ∆ a T , which ensures that a certainadvantage is achieved through the prospected lane-changemaneuver. This is formalized by the following inequality: ∆ a c + p (∆ a n + ∆ a o ) > ∆ a T . (9)The parameter values that we used for both IDM andMOBIL controllers are detailed in Table III. Remarks.
Testing the aforementioned models on ourexperimental testbed demonstrated the need for adaptationsthat take into account the physical vehicles’ kinematic anddynamic constraints.Firstly, we noticed the the jam distance s requires anadditional escape distance, which is a function of the egovehicle’s desired speed and the front vehicle’s current speed, v f , in order facilitate lane merging. Hence, we used aneffective jam distance ˆ s = s + s e , where s e follows therule: s e ( v f , v ) =
0, if v f /v > L (cid:20) (cid:16) v f v (cid:17) − (cid:16) v f v (cid:17) + 1 (cid:21) , else.This ensures a larger escape distance when the speed ofthe front vehicle is small relative to the ego car’s desiredspeed, and that this distance approaches zero as the speedsbecome equal. Secondly, a lane change on real cars is notimmediate, and takes time to be completed. Hence, duringthis transition, cars on the original lane will still consider thestate of ego vehicle in their longitudinal control, until theego vehicle has completely entered the new lane. Similarly,the ego car starts using information about the state of newpreceding and following cars on the new lane, as soon asit starts changing lanes. Finally, an extra safety constraint isadded to the MOBIL model to check that the escape distancebetween the ego vehicle and the front vehicle is large enoughto prevent crashing when changing lanes. B. Cooperative Driving
Our cooperative driving strategy builds on the assumptionthat vehicles within visibility range c communicate withne another to share intended maneuvers before actuallyexecuting them. This allows the vehicles to cooperate aboutlane-changing decisions, and hence, plan efficient paths thatmaximize traffic throughput, whilst ensuring safety. To testour setup’s capability of validating the effects of coopera-tive driving we implement an approach that builds on thefollowing key ideas: • When the ego vehicle decides to change its lane, itcommunicates with its neighboring vehicles, projectinga virtual counterpart vehicle at the desired new state. • Vehicles within communication range of the ego vehiclereceive information about its projected (virtual) state.They take this information into account to accelerate ordecelerate, as a function of their relative positions to thevirtual vehicle.We implement this behavior by modifying the original IDMand MOBIL models as follows.
Cooperative -IDM (C-IDM).
Our cooperative-IDM modeltakes projected vehicles into account in the form of weightedvirtual vehicles. When a vehicle projects its desired state, theresulting virtual vehicle is given a weight w v between 0 and1 depending on how urgent the lane-change is. This urgencyis defined by w v = min(1 , κ ( c − s )) , where s is the gapbetween the actual ego vehicle (to whom the virtual vehiclebelongs), and its preceding vehicle. Based on this idea, ourcooperative IDM model differs from the original IDM modelin two ways. First, our model decelerates when detectinga projected vehicle in front of it. It does this by returningan acceleration a = min( w v · ˜ a IDM , a
IDM ) , where a IDM isthe acceleration computed using an actual front vehicle (ifpresent), and where ˜ a IDM is the acceleration computed usingthe projected front vehicle. Second, it increases the desiredspeed when in front of or too close to a virtual vehicle,so that it can make space for the desired, projected vehiclestate of the ego vehicle that owns the virtual vehicle. This isformalized by the following rule: ˜ v = v (cid:18) w v c − ˜ s trail c (cid:19) where ˜ s trail is the distance between the neighboring vehicleand the virtual vehicle behind it. Cooperative -MOBIL (C-MOBIL).
Our cooperative-MOBIL model uses a higher safe braking acceleration valuethan the original model in order to pack vehicles more closelywhen changing lanes; we set β n = α . Also, we add a safetyconstraint to prevent crashing when the speeds are low andthe acceleration given by IDM is low: s > s + γ ∆ v , where s is the gap between the ego vehicle and the front/rear vehicleon the nearby lane. The constant γ corresponds to the timeneeded for an average lane change. C. Gamification
We gamify our experimental setup by interfacing one car(or several cars) with a joystick or keyboard. This allows ahuman player to experience traffic amid different types ofsurrounding vehicle behaviors. It also allows us to stress-testthe reactions of other cars in traffic to arbitrary maneuvers.A user can choose between three modes of control:‘manual’, ‘semi-automatic’ and ‘automatic’. ‘Manual’ givesPSfrag replacements
X [m] Y [ m ] start ReferenceReferenceMinicarMinicar
Fig. 6:
Overhead representation of trajectory tracking accuracy for one loopof our U-shaped reference trajectory. a user direct control over the speed and turning angle of thecar. ‘Semi-automatic’ confines the vehicle to the designatedlanes and safety restrictions while allowing the user to selecta maintained speed and change lanes. ‘Automatic’ removesall control from the player and integrates the vehicle into thetraffic. VII. E
XPERIMENTS
We perform three sets of experiments on our vehicle fleet.Our aim is to (a) demonstrate the navigation capabilities ofour Minicar, (b) demonstrate the operation of the Minicarfleet, and (c) show how the fleet is used to test and validate(novel) driving algorithms in a realistic, albeit miniature,setup.
A. Trajectory Tracking
We validate the Minicar’s trajectory tracking capabilities.We run the Minicar on our U-shaped reference trajectoryat a nominal speed of 0.4m/s. Figure 6 shows an overheadplot of the tracking accuracy, for one loop on our U-shapedtrack. The average tracking error, measured over 7 loops ofour track, is 14 mm with a standard deviation of 6.3 mm.
B. Driving Behaviors in Multi-Car Traffic
We implement four schemes to show the effect of differ-ent driving behaviors in multi-car traffic. We consider twodriving policies: all cars are driven by (1) the egocentricpolicy (described in Sec. VI-A), or (2) , the cooperative policy(described in Sec. VI-B) with c = 2 m. For each policy,we use either the normal or the aggressive parameter set(see Table III). An experiment involves 16 Minicars drivingon our U-shaped track and runs for 200s. At the start, theMinicars are evenly spaced out over the track. After 20s, oneof the cars is told to stop (and hence, it blocks traffic on itscurrent lane). The aim is to observe how the four differentschemes react to this disturbance.Figure 7 shows the traffic flow for these four experiments,on the top row for the egocentric driving policy, and onthe bottom row for the cooperative driving policy. In theegocentric scheme, we observe how the blockage creates avehicle queue, which increases to 5 waiting vehicles at thelatest stage. In contrast, the cooperative scheme overcomesvehicle queuing altogether (the traffic patterns exhibit veryshort stationary phases). With cooperative behavior, instead .36.69.913.216.50.0 3.36.69.913.216.50.0 PSfrag replacements P o s iti on a l ong T r ac k [ m ] P o s iti on a l ong T r ac k [ m ] Time [s]Time [s]
Egocentric (Aggressive)Egocentric (Normal)
Cooperative (Aggressive)Cooperative (Normal) V e l o c it y [ m / s ] PSfrag replacements P o s iti on a l ong T r ac k [ m ] P o s iti on a l ong T r ac k [ m ] Time [s]Time [s]
Egocentric (Aggressive)Egocentric (Normal)
Cooperative (Aggressive)Cooperative (Normal) V e l o c it y [ m / s ] Fig. 7:
Results for experiments performed on 16 Minicars driving on our multi-lane U-shaped track, for egocentric (top) and cooperative (bottom) drivingpolicies. The panels show vehicle positions the track, plotted as a function of time. One car is stopped after 20s to cause a traffic disturbance. The colorbarshows the velocity recorded. The panels on the left show data for the normal parameter settings, panels on the right show the aggressive parameter settings(see Table III).
PSfrag replacements P o s iti on a l ong T r ac k [ m ] V e l o c it y [ m / s ] Time [s]
Accel.TurnBrake brakeaccel.leftrightstop
Fig. 8:
A gamified experiment where a human controls one Minicar viaa joystick, and 10 other Minicars are automated. The thick trajectoryrepresents the played car. The colorbars along the time axis show the periodsduring which control commands are executed.
Egocentric Cooperative
Improvement
Normal . ± .
036 0 . ± . Aggressive . ± .
045 0 . ± . TABLE IV:
Average throughput and standard deviation, measured incars-per-second, for the four experiments shown in Figure 7. The thirdcolumn shows the percent improvement of the cooperative scheme overthe egocentric scheme. of queuing, a Minicar communicates its intention to lane-change; following vehicles in the new lane reduce theirspeeds to make space for this projected maneuver, hencemaintaining traffic flow whilst ensuring safety. The overallthroughput is improved in the cooperative scheme, andis further enhanced by aggressive control parameters; theaverage throughput values are reported in Table IV.
C. Gamification
Figure 8 shows an example of a player-controlled car amid10 automated cars, in cooperative mode. We observe how thetraffic is affected by the actions of the human player.VIII. D
ISCUSSION
In this work, we provided the design of a fleet of miniaturecars for research, education and outreach in the domain ofautomated multi-car systems. Our Ackermann-steering plat-form is one out of very few openly available designs; in par-ticular, it fills a price-range gap, and is especially attractivefor robotics labs that already possess telemetry infrastructure(such as motion capture). We propose a system architecturethat is capable of integrating heterogeneous driving strategies(e.g., egocentric or cooperative) in a multi-lane setup. Wedemonstrate its applicability for large-fleet experimentationby implementing four different driving schemes that lead toquantifiable, distinct traffic behaviors.Although our current setup considers off-board intelli-gence, the platform is easily extended with an IMU andcamera to provide full on-board autonomy. In future work,we plan to use our fleet for testing multi-car systems inmore complex scenarios that include: (a) road topographieswith a larger number of lanes as well as intersections, (b) heterogeneous vehicle behaviors in mixed traffic, and (c) noisy sensing and delayed communications. Further researchwill investigate multi-objective optimization problems thatalso include driver comfort (as measured by vehicle acceler-ations).
EFERENCES [1] CBS Insights Research Brief. .(Accessed August 15, 2018).[2] Hemisson. . (Accessed August 13, 2018).[3] Sribbler 3. . (Accessed August 13, 2018).[4] M. Dekan, F. Duchon, et al. irobot create used in education.
Journal of Mechanics Engineering and Automation , 3(4):197–202, 2013.[5] F. Dressler, H. Hartenstein, O. Altintas, and O. Tonguz. Inter-vehicle communication: Quo vadis.
IEEE CommunicationsMagazine , 52(6):170–177, 2014.[6] M. Ferreira, R. Fernandes, H. Conceic¸ ˜ao, W. Viriyasitavat, andO. K. Tonguz. Self-organized traffic control. In
Proceedingsof the seventh ACM international workshop on VehiculArInterNETworking , pages 85–90. ACM, 2010.[7] J. Gonzales, F. Zhang, K. Li, and F. Borrelli. Autonomousdrifting with onboard sensors. In
Advanced Vehicle Con-trol: Proceedings of the 13th International Symposium onAdvanced Vehicle Control (AVEC16), September 13-16, 2016,Munich, Germany , page 133, 2016.[8] A. Jim´enez-Gonz´alez, J. R. Martinez-de Dios, and A. Ollero.Testbeds for ubiquitous robotics: A survey.
Robotics andAutonomous Systems , 61(12):1487–1501, 2013.[9] S. Karaman, A. Anders, M. Boulet, J. Connor, K. Gregson,W. Guerra, O. Guldner, M. Mohamoud, B. Plancher, R. Shin,et al. Project-based, collaborative, algorithmic robotics forhigh school students: Programming self-driving race cars atMIT. In
Integrated STEM Education Conference (ISEC), 2017IEEE , pages 195–203. IEEE, 2017.[10] S. Kernbach. Swarmrobot. org-open-hardware microroboticproject for large-scale artificial swarms. arXiv preprintarXiv:1110.5762 , 2011.[11] A. Kesting, M. Treiber, and D. Helbing. General lane-changing model MOBIL for car-following models.
Trans-portation Research Record , 1999(1):86–94, 2007.[12] E. King, Y. Kuwata, M. Alighanbari, L. Bertuccelli, andJ. How. Coordination and control experiments on a multi-vehicle testbed. In
American Control Conference, 2004.Proceedings of the 2004 , volume 6, pages 5315–5320. IEEE,2004.[13] M. Linderoth, K. Soltesz, and R. M. Murray. Nonlinear lateralcontrol strategy for nonholonomic vehicles. In
AmericanControl Conference, 2008 , pages 3219–3224. IEEE, 2008.[14] A. Liniger, A. Domahidi, and M. Morari. Optimization-based autonomous racing of 1: 43 scale rc cars.
Optimal ControlApplications and Methods , 36(5):628–647, 2015.[15] J. McLurkin, A. McMullen, N. Robbins, G. Habibi, A. Becker,A. Chou, H. Li, M. John, N. Okeke, J. Rykowski, et al. Arobot system design for low-cost multi-robot manipulation. In
Intelligent Robots and Systems (IROS 2014), 2014 IEEE/RSJInternational Conference on , pages 912–918. IEEE, 2014.[16] L. Paull, J. Tani, H. Ahn, J. Alonso-Mora, L. Carlone, M. Cap,Y. F. Chen, C. Choi, J. Dusek, Y. Fang, et al. Duckietown:an open, inexpensive and flexible platform for autonomyeducation and research. In
Robotics and Automation (ICRA),2017 IEEE International Conference on , pages 1497–1504.IEEE, 2017.[17] D. Pickem, P. Glotfelter, L. Wang, M. Mote, A. Ames,E. Feron, and M. Egerstedt. The robotarium: A remotelyaccessible swarm robotics research testbed. In
Robotics andAutomation (ICRA), 2017 IEEE International Conference on ,pages 1699–1706. IEEE, 2017.[18] J. A. Preiss*, W. H¨onig*, G. S. Sukhatme, and N. Aya-nian. Crazyswarm: A large nano-quadcopter swarm. In
IEEE International Conference on Robotics and Automation(ICRA) , pages 3299–3304. IEEE, 2017. Software available at https://github.com/USC-ACTLab/crazyswarm .[19] F. Riedo, M. Chevalier, S. Magnenat, and F. Mondada.Thymio II, a robot that grows wiser with children. In , pages 187–193. Eidgen¨ossische TechnischeHochschule Z¨urich, Autonomous System Lab, 2013.[20] M. Rubenstein, C. Ahler, and R. Nagpal. Kilobot: A low costscalable robot system for collective behaviors. In
Roboticsand Automation (ICRA), 2012 IEEE International Conferenceon , pages 3293–3298. IEEE, 2012.[21] M. Rubenstein, B. Cimino, R. Nagpal, and J. Werfel. Aerobot:An affordable one-robot-per-student system for early roboticseducation. In
Robotics and Automation (ICRA), 2015 IEEEInternational Conference on , pages 6107–6113. IEEE, 2015.[22] M. Treiber, A. Hennecke, and D. Helbing. Congested trafficstates in empirical observations and microscopic simulations.
Physical review E , 62(2):1805, 2000.[23] G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A.Theodorou. Aggressive driving with model predictive pathintegral control. In
Robotics and Automation (ICRA), 2016IEEE International Conference on , pages 1433–1440. IEEE,2016.[24] S. Wilson, R. Gameros, M. Sheely, M. Lin, K. Dover,R. Gevorkyan, M. Haberland, A. Bertozzi, and S. Berman.Pheeno, a versatile swarm robotic research and educationplatform.