This AI Learned to Drive Like a Human With 2,500x Less Data Than Normal
New method uses just 30 minutes of human driving data to train AI that drives like humans.
A team including Daphne Cornelisse (Cornell) and Eugene Vinitsky (Princeton) has published a paper demonstrating a breakthrough in autonomous driving AI that achieves human-like behavior with surprisingly little human input. Their method leverages self-play reinforcement learning (where AI agents learn by interacting with each other) while using just 30 minutes of human driving demonstrations as a 'regularization objective' to guide behavior toward human-like conventions.
The approach solves a key challenge in pure self-play systems, where AI agents develop alien driving habits incompatible with human road users. Unlike prior attempts that require extensive reward engineering or massive human datasets, this method achieved coordination with held-out human trajectories using 2,500 times less human data than typical imitation learning approaches. Training completes in just 15 hours on a single consumer-grade GPU, making it far more accessible than conventional training methods that require massive computational resources.
- Uses 30 minutes of human data (2,500x less than comparable methods) to guide AI behavior
- Combines self-play reinforcement learning with minimal human demonstrations for human-like autonomy
- Trains in 15 hours on a single consumer GPU - far faster than traditional approaches
Why It Matters
This could dramatically reduce the cost and complexity of training human-like autonomous systems for robotics, self-driving vehicles, and human-AI collaboration scenarios.