Hey, I'm jacques

Current project

Backseat driving: Using privileged-information auxiliary losses in autonomous racing

End-to-end reinforcement learning (RL) agents can drive a car directly from raw sensor inputs, but a LiDAR-only agent must implicitly learning quantities it cannot observe, such as its own velocity, tyre slip, and the curvature of the track ahead to perform well. In simulation, these quantities are available as privileged information, yet standard RL provides no direct signal to learn them; even when the reward is dense, it is still an indirect and high-variance teaching signal for these quantities, so the agent learns them slowly. This project investigates whether auxiliary losses that predict privileged information can be applied to improve the performance and training speed of a LiDAR- only end-to-end racing agent, and whether they change what the agent’s internal representations encode. We find that close-ahead track curvature prediction has a positive effect on the training speed, racing performance, crash rate, and representation quality of the agent, while other losses such as velocity and slip prediction have little-to-no effect.