Reproduce “FastRLAP: A System for Learning High-Speed Driving via Deep RL and Autonomous Practicing” on OpenResearch

About this paper

Kyle Stachowicz, Dhruv Shah, Arjun Bhorkar, Ilya Kostrikov, Sergey Levine

FastRLAP combines navigation-specific visual pretraining, a single slow reconnaissance lap, and autonomous online reinforcement learning to teach a small RC car high-speed driving on previously unseen courses. Across indoor, outdoor, and simulated tracks, the system learned faster racing lines, braking, drifting, and terrain selection without a global map or repeated human resets, reducing example lap times from 54–160 seconds to 10.9–62.1 seconds in the reported real-world settings.

Key insights

• Prior navigation experience is useful primarily as a transferable visual representation, not as a ready-made driving policy; online interaction remains necessary to discover aggressive behavior suited to the new course.
• Simple recovery behaviors can substitute for full physical resets, allowing continuous practice even when the vehicle collides or becomes stuck.
• The learned policy can exploit visual and terrain cues to develop behaviors absent from the slow demonstration, including late braking, post-corner acceleration, drifting, racing-line selection, and avoidance of low-traction grass or mud.

Summary by alphaXiv.