Practical implementation of obstacle avoidance strategies in the truly multi-vehicle Autonomous Robot Racing Competition

This paper introduces a practical approach to handle object tracking and path planning methodology for real-world multi-vehicle autonomous racing, interacting with more than 8 vehicles. Unlike previous autonomous racing systems, which primarily dealt with single or dual races, our proposed algorithm successfully handles real-world multi-vehicle scenarios, demonstrated in the “Autonomous Robot Racing Competition”(ARRC) with nine vehicles. The perception module utilizes a 16-channel LiDAR sensor to detect multiple obstacles on the racing track. To overcome the challenges posed by sparse point clouds, we introduce an orientation compensation method of multi-object detection on sparse point cloud conditions by applying the Extended Kalman Filter(EKF) tracking method. Our algorithm demonstrated 99.6% of the overall orientation accuracy compared to learning based methods that use 64-channel or higher resolution LiDAR. Moreover, it performed better when recognizing small objects with fewer points. The behavior predictive motion planning algorithm predicts dynamic multiple opponents’ trajectories and generates candidate paths considering two racing lanes and the states of other multiple vehicles applying the Frenet-Frame. The proposed algorithm is tested in a custom CARLA simulator for 20 scenarios with multi-vehicle interaction, and its effectiveness is demonstrated in the real-world 2023 ARRC. Our algorithm achieves safe overtaking, avoidance, and following maneuvers through multi-vehicle racing while adhering to the given racing rules.