Jack Collier

Learning-based Nonlinear Model Predictive Control to Improve Vision-based Mobile Robot Path Tracking

This paper presents a Learning-based Nonlinear Model Predictive Control (LB-NMPC) algorithm for an autonomous mobile robot to reduce path-tracking errors over repeated traverses along a reference path. The LB-NMPC algorithm uses a simple a priori vehicle model and a learned disturbance model. Disturbances are modelled as a Gaussian Process (GP) based on experience collected during previous traversals as a function of system state, input and other relevant variables. Modelling the disturbance as a GP enables interpolation and extrapolation of learned disturbances, a key feature of this algorithm. Localization for the controller is provided by an on-board, vision-based mapping and navigation system enabling operation in large-scale, GPS-denied environments. The paper presents experimental results including over 1.8 km of travel by a four-wheeled, 50 kg robot travelling through challenging terrain (including steep, uneven hills) and by a six-wheeled, 160 kg robot learning disturbances caused by unmodelled dynamics at speeds ranging from 0.35 m/s to 1.0 m/s. The speed is scheduled to balance trial time, path-tracking errors, and localization reliability based on previous experience. The results show that the system can start from a generic a priori vehicle model and subsequently learn to reduce vehicle- and trajectory-specific path-tracking errors based on experience.

Software Systems for Robotics An Applied Research Perspective

Over the past 20 years, Defence Research and Development Canada has developed numerous tele-operated unmanned ground vehicles (UGV), many founded on the ANCÆUS command and control system. This paper relates how long experience with tele-operated UGVs influenced DRDCs shift in focus from tele-operated to autonomous unmanned vehicles (UV), the forces that guided DRDCs development approach and DRDCs experience adapting a specific tool set, MIRO, to a UGV implementation.

Continuous motion, outdoor, 2 1/2D grid map generation using an inexpensive nodding 2-D laser rangefinder

This paper introduces a technique for creating 2 1/2D grid maps of unstructured, outdoor environments, while traveling at high speeds, using an inexpensive nodding 2-D laser rangefinder. The nodding mechanism allows the acquisition of multiple range data sets for terrain in front of the robot. While these multiple data sets alleviate some of the problems traditionally associated with laser rangefinders, they also introduce a new set of problems. The paper investigates and quantifies factors that determine the accuracy of a map generated using a nodding laser rangefinder and derives an optimal basis for minimizing these errors. This research has determined that the most significant source of errors, for a nodding laser rangefinder configuration, are the roll, pitch and yaw accuracy for the laser beam. A variance weighted statistical approach was implemented to optimally fuse the range data into the 2 1/2D grid map. Simulations and experiments were conducted, demonstrating the performance of the variance weighted technique as superior to classical statistical methods

Architecture for autonomy

In 2002 Defence R&D Canada changed research direction from pure tele-operated land vehicles to general autonomy for land, air, and sea craft. The unique constraints of the military environment coupled with the complexity of autonomous systems drove DRDC to carefully plan a research and development infrastructure that would provide state of the art tools without restricting research scope. DRDCs long term objectives for its autonomy program address disparate unmanned ground vehicle (UGV), unattended ground sensor (UGS), air (UAV), and subsea and surface (UUV and USV) vehicles operating together with minimal human oversight. Individually, these systems will range in complexity from simple reconnaissance mini-UAVs streaming video to sophisticated autonomous combat UGVs exploiting embedded and remote sensing. Together, these systems can provide low risk, long endurance, battlefield services assuming they can communicate and cooperate with manned and unmanned systems. A key enabling technology for this new research is a software architecture capable of meeting both DRDCs current and future requirements. DRDC built upon recent advances in the computing science field while developing its software architecture know as the Architecture for Autonomy (AFA). Although a well established practice in computing science, frameworks have only recently entered common use by unmanned vehicles. For industry and government, the complexity, cost, and time to re-implement stable systems often exceeds the perceived benefits of adopting a modern software infrastructure. Thus, most persevere with legacy software, adapting and modifying software when and wherever possible or necessary -- adopting strategic software frameworks only when no justifiable legacy exists. Conversely, academic programs with short one or two year projects frequently exploit strategic software frameworks but with little enduring impact. The open-source movement radically changes this picture. Academic frameworks, open to public scrutiny and modification, now rival commercial frameworks in both quality and economic impact. Further, industry now realizes that open source frameworks can reduce cost and risk of systems engineering. This paper describes the Architecture for Autonomy implemented by DRDC and how this architecture meets DRDCs current needs. It also presents an argument for why this architecture should also satisfy DRDCs future requirements as well.

Speed Daemon: Experience-Based Mobile Robot Speed Scheduling

A time-optimal speed schedule results in a mobile robot driving along a planned path at or near the limits of the robots capability. However, deriving models to predict the effect of increased speed can be very difficult. In this paper, we present a speed scheduler that uses previous experience, instead of complex models, to generate time-optimal speed schedules. The algorithm is designed for a vision-based, path-repeating mobile robot and uses experience to ensure reliable localization, low path-tracking errors, and realizable control inputs while maximizing the speed along the path. To our knowledge, this is the first speed scheduler to incorporate experience from previous path traversals in order to address system constraints. The proposed speed scheduler was tested in over 4 km of path traversals in outdoor terrain using a large Ackermann-steered robot travelling between 0.5 m/s and 2.0 m/s. The approach to speed scheduling is shown to generate fast speed schedules while remaining within the limits of the robots capability.

Integration of a high degree of freedom robotic manipulator on a large unmanned ground vehicle

The Multi-Agent Tactical Sentry Unmanned Ground Vehicle, developed at Defence R&D Canada - Suffield, has been in service with the Canadian Forces for five years. This tele-operated wheeled vehicle provides a capability for point detection of chemical, biological, radiological, and nuclear agents. Based on user experience, it is obvious that a manipulator capability would greatly enhance the vehicles utility and increase its mobility in urban terrain. This paper details technical components of this development, and describes a number of trials undertaken to perform tasks with a manipulator arm such as picking up objects, opening vehicle and building doors, recording video, and creating 3D models of the environment. The lessons learned from these trials will guide further development of the technology.

Environment Classification for Indoor/Outdoor Robotic Mapping

We present a novel perception system for mapping of indoor/outdoor environments with an Unmanned GroundVehicle (UGV). The system uses image classification techniques to determine the operational environment of theUGV (indoor or outdoor). Based on the classification results, the appropriate mapping system is then deployed.Image features are extracted from video imagery andused to train a classification function using supervisedlearning techniques. This classification function is thenused to classify new imagery. A perception module observesthe classification results and switches the UGVs perception system, according to current needs and available (reliable) data as the UGV transitions from indoors to outdoors or vice versa. A terrain map that exploits GPS and Inertial Measurement Unit (IMU) data is used when operatingoutdoors, while a 2D laser based Simultaneous Localization and Mapping (SLAM) technique is used when operating indoors. Globally consistent maps are generated bytransforming the indoor map data into the global referenceframe, a capability unique to this algorithm.

Multi-sensor Appearance-Based Place Recognition

This paper describes a multi-sensor appearance-based place recognition system suitable for robotic mapping. Unlike systems that extract features from visual imagery only, here we apply the well known Bag-of-Words approach to features extracted from both visual and range sensors. By applying this technique to both sensor streams simultaneously we can overcome the deficiencies of each individual sensor. We show that LIDAR-based place recognition using a generative model learnt from Variable Dimensional Local Shape Descriptors can be used to perform place recognition regardless of lighting conditions or large changes in orientation, including traversing loops backward. Likewise, we are still able to exploit the feature rich place recognition that visual systems provide. Using a pose verification system we are able to effectively discard false positive loop detections. We present experimental results that highlight the strength of our approach and investigate alternative techniques for combining the results from the individual sensor streams. The multi-sensor approach enables the two sensors to complement each other well in large urban and rural environments under variable lighting conditions.

The robotics experience

The classical engineering fields have evolved standards and techniques for developing complex systems. For example, both mechanical and electrical engineers have a wide variety of standard components, with defined capabilities, that they can draw upon (e.g., gears, transistors) in the design of complex systems. On the other hand, software engineering has struggled with the basic idea of reusability. Software engineering approaches, such as the use of components that promote the concept of information hiding and the introduction of structured programming languages, offer a roadmap to an improved software reuse. Unfortunately, their adoption by robotics researchers has been slow, impeded by the tradition of individual research groups crafting independent and incompatible solutions to common problems.

Probabilistic Obstacle Detection Using 2 1/2 D Terrain Maps

Navigating unstructured environments requires reliable perception that generates an appropriate world representation. This representation must encompass all types of impediments to traversal, whether they be insurmountable obstacles, or mobility inhibitors such as soft soil. Traditionally, traversability and obstacle avoidance have represented separate capabilities with individual rangefinders dedicated to each task. This paper presents a statistical technique that, through the analysis of the underlying 21/2 D terrain map, determines the probability of an obstacle. This integrated approach eliminates the need for multiple data sources and is applicable to range data from various sources, including laser rangefinders and stereo vision. The proposed obstacle detection technique has been tested in simulated environments and under real world conditions, and these experiments revealed that it accurately identifies obstacles.