DescriptionIn the future of the autonomous car industry saving lives might depend on more demanding maneuvers than what the average drivers know how to do. Professional race car drivers, as well as some skilled stunt drivers, are able to exteremely push the limits of a vehicle's capabilities and safety features which are useful to avoid hazardous situations. By understanding these human-inspired driving abilities, a motion planning and autonomous vehicle control system can be developed for enabling Aggressive Maneuvering as a Safety Feature (AMSF). Similar to existing and widely used vehicle safety features, such as anti-lock braking and electronic stability control, AMSF can be utilized to increase overall safety. AMSF system will push the limits of maneuverability and motion stability to enable the next generation of accident-free vehicle systems. The design of an AMSF requires advanced theoretical tools and algorithms to guarantee safety assurance under a dynamically changing environment. Current AMSF design methods incorporate a variety of elements such as analytical vehicle dynamics model-based control, machine learning-based methods to mimic expert human drivers, integrated physical model-based knowledge, and experience-based skills to enhance vehicle maneuverability with guaranteed motion stability.
The main goal of this dissertation focuses on the safety-guaranteed motion planner and controller design for AMSF. This dissertation proposes methods and algorithms that enable the integration of data-driven approaches based on machine learning techniques and physical model analysis that will achieve parity with and even exceed the driving skills of the most skilled drivers. To achieve these goals, several new modeling and control approaches are developed in this dissertation. First, in order to improve the physical model accuracy under a dynamic environment, Gaussian Process on Polynomial Basis (GPPB) method is proposed to learn from human expert driving data. A Sum Of Square (SOS) method is used to estimate the safety boundary of the nonlinear vehicle dynamics to enable the use of Nonlinear Model Predictive Controller (NMPC) for AMSF design. It is also shown that the safety of the aggressive maneuvering can be considered as a safety brier applied as constraints to the vehicle's motion similar to a control barrier function. The main rationale of these designs comes from observations that many physical models cannot effectively capture dynamic changes or the uncertainties of the vehicle/environment interactions, such as tire-road contact properties. Machine learning-based methods generally provide an effective means to obtain the system dynamics changes in real time and incorporate them into the control design.
One important feature of the above integrated GPPB and SOS approach lies in the guaranteed safety and stability under the control design for aggressive vehicle maneuvers. The motion safety assurance while tracking a planned trajectory has been successfully resolved by integrating the data-driven methods with physical model-based controllers under dynamic changing environments. Theoretical analyses are presented in the dissertation for the vehicle dynamics model and also extendable to other well-understood dynamics structure of the car-like robots. The Lyapunov stability method is used to show safety assurance of a learning model-based control framework. A scaled race car-like robot is used as an experimental testbed to demonstrate the proposed control of the AMSF design. The experimental results demonstrate superior agility and fast traveling time performance under the proposed control design than those under the physical model-based control design in literature.
Another important aspect of this dissertation focuses on the real-time motion planning for AMSF design. Planning the trajectory for an autonomous vehicle for agile maneuvers in a dynamic environment is complex and challenging, especially when the autonomous vehicles tend to maximize the driving capabilities to achieve certain metrics such as traveling time, etc. The proposed motion planner in the dissertation takes advantage of the Sparse Stable Trees (SST), the Star Rapidly-exploring Random Tree (RRT*) algorithm, and the NMPC design. The use of the sparsity property helps to reduce the computational burden of the RRT* method by removing non-used nodes in each iteration and therefore to render the algorithm to converge to optimal path quickly. A heuristic quality function is used to guide the search to achieve faster convergence, and the NMPC is used for rewiring feasibility among nodes. The motion planner is tested experimentally and also compared with the existing benchmarks to demonstrate superior performance.
Finally, the last part of the dissertation discusses and proposes an extension of the above-mentioned methods to further incorporate the newly development of machine learning techniques. Another goal is to test Gaussian Process and Reinforcement Learnings for direct and indirect controller design. Our goal is to test the learning methods for both the off-line and the online applications and design stable adaptive rules for real-time stability and safety assurance and performance enhancement. Taking advantage of policy search approaches alongside nonlinear programming methods gives strong optimization power for overall safety and agility in real-time implementation. The works in this part will be experimentally validated and demonstrated using two different car-like robot platforms: one is racing car-like and the other is a scaled autonomous truck used for the minimum time lap and stunt maneuverings experiments.