Towards fully autonomous racing

Abstract

Autonomous racing is a new field of research that is gaining a lot of attention. Racing presents numerous research and development challenges due to the fact that a vehicle is driven at its perception, planning, and control limits. While the majority of autonomous driving research has focused on routine driving situations, achieving the safety and performance improvements of autonomous vehicles also requires a focus on driving under extreme conditions. This dissertation’s primary focus can be divided into three folds: 1. Design of autonomy system for the Indy Autonomous Challenge. 2. Imitation learning approach using hierarchical policy abstractions level for the 1:n player race. 3. Lessons and open research questions in the field of autonomous racing. Our autonomous racing system mainly comprises perception, planning, control, and system status monitoring subsystems. Along with its performance, various critical aspects of field robotics (such as real-time computation capability, resiliency, and fault tolerance) are taken into account from a real-world deployment point of view. The proposed system was integrated into a full-scaled autonomous race car (Dallara AV-21) and extensively validated during field testing. Our system successfully performed all missions including overtaking at speeds over 200 kph in the CES 2022 Indy Autonomous Challenge(IAC), the world’s first head-to-head autonomous race. Towards the fully autonomous race, which is analogous to races between professional human drivers, we proposed the imitation learning approach using hierarchical policy abstractions for planning and control. In the trajectory-level abstraction, our policy model outputs the imitation prior distribution, which indicates the likelihood of the expert’s future trajectory. Then, the model selects the maximum likelihood trajectory given a set of candidate trajectories and passes it to the control policy, trained to generate the control signal in a supervised fashion. The proposed imitation learning method is fully integrated with the aforementioned autonomous racing stack and evaluated using a high-fidelity racing simulator containing various baselines. The quantitative study results show that our method outperforms the other baseline methods in terms of planning accuracy. Further, closed loop simulation results show that our method can compete with other agents under challenging head-to-head race scenarios. Last but not least, this dissertation delivers lessons and open research questions that we learned from the unique field experience using high-speed, full-scaled autonomous race vehicles. These are derived from the autonomous racing application but can also be applied to relevant robotics research providing future direction and insight.