Human-like autonomy emerges from self-play and a pinch of human data

Short summary

A new self-play reinforcement learning method trains autonomous driving policies using only 30 minutes of human demonstrations—2500x less than traditional imitation learning. The approach treats human data as regularization atop a minimal safe reward signal, learning effective driving behavior in 15 hours on consumer GPUs. Code and results are publicly available.

• •Uses 30 minutes of human data vs. traditional imitation learning's 2500x more data
• •Trains in 15 hours on consumer-grade GPUs with self-play reinforcement learning
• •Policies coordinate with human trajectories; full source code available

Generated with AI, which can make mistakes.