Abstract
Autonomous racing is an emerging research area that presents unique, unsolved, and challenging problems. Racing involves operating vehicles at the limits of their perception, planning, and control capabilities. To succeed in racing, autonomous agents must predict the future motion of other racing agents without explicit signals or lane boundaries, then plan maneuvers that overtake opponent vehicles while pushing the vehicle to its kinematic and dynamic limits, all without leaving the drivable area defined by the track boundaries. The adversarial nature of racing, coupled with its unstructured environments and minimal margin for error make autonomous racing a rich problem space that existing autonomous vehicle technologies do not fully address. There is rapidly growing interest in this space. Numerous autonomous racing platforms have been developed in recent years, including at the laboratory scale (such as F1/10) as well as full-scale autonomous racing competitions like A2RL and the Indy Autonomous Challenge (IAC), which features head-to-head racing at speeds up to 180MPH.
However, key challenges in motion planning for autonomous racing remain unsolved. Many existing works either lack explicit guarantees about the racing agent’s behavior or require overly simplistic assumptions that make them unsuitable for complex road courses. We consider opponent trajectory prediction, collision risk estimation, and trajectory synthesis in high-speed racing scenarios and present several contributions towards addressing these unsolved challenges.
First, we present the DeepRacing simulation and testbed, the first full-scale racing testbed of its kind. DeepRacing builds on decades of physics and graphics research from the video games industry by leveraging the F1© video game series from Codemasters™. DeepRacing allows collection of realistic data in complex multi-agent racing scenarios as well as closing the loop and testing racing algorithms on Formula One racetracks.
Second, we explore end-to-end methods for autonomous racing with data collected from DeepRacing and present a neural network based approach to predict Bézier curves as a low-dimensional representation of a reference trajectory. This technique outperforms several methods that directly prediction control actions as well as existing methods in trajectory synthesis.
We then present a Bayesian interpretation of trajectory synthesis called Differential Bayesian Filtering (DBF) designed to address key limitations of end-to-end methods for trajectory synthesis by incorporating knowledge of the vehicle’s dynamic limits with probabilistic inference. We show that DBF exhibits superior performance on several racing metrics in a single-agent setting, even outperforming all of the human data it was trained on.
To enable multi-agent extensions of DBF, we next present BARTé: a neural-network based technique for predicting future motion of opponent vehicles. BARTé addresses limitations of existing methods for racing trajectory prediction by applying key mathematical properties of composite Bézier curves. BARTé outperforms existing methods, both those for urban driving and state of the art methods in racing, in terms of accuracy and computational efficiency while guaranteeing C1 continuity in its predictions.
However, a trajectory prediction of opponent vehicle is not sufficient per se for multi- agent racing. To enable collision avoidance in our probabilistic DBF framework, we present the Gauss-Legendre Rectangle (GLR) algorithm for estimating the cumulative probability of collision between a prediction opponent trajectory and a candidate ego reference trajectory. Derived from Poisson process theory and Gauss-Legendre quadrature, GLR does not make require simplifying assumptions common in existing methods for cumulative risk estimation and outperforms existing methods while maintaining the fast computation time required in racing applications.
Finally, we bring the DBF framework, trajectory prediction, and collision risk estimation together into a trajectory synthesis method for multi-agent autonomous racing called DBF-MA. This method synthesizes overtaking maneuvers with probabilistic inference over the space of composite Bézier curves. DBF-MA successfully plans an overtake maneuver in over 87% of closed-loop scenarios and offers explicit guarantees on continuity and a smooth transition back to the globally optimal trajectory. We show that this method outperforms several baseline methods in racing trajectory synthesis in terms of collision avoidance and dynamic feasibility, all while maintaining a 10Hz loop rate.
