Autonomous driving AI trained two ways at once instead of one after another — achieves better results

Researchers built a new training method that teaches self-driving cars using human demonstrations and reinforcement learning simultaneously, instead of learning from humans first and then fine-tuning with trial-and-error. In practice, this means the AI can reach better performance without getting stuck repeating human mistakes, and retraining one component doesn't break the other.

The standard approach to autonomous driving has been a bottleneck: you train on human driving videos until the AI plateaus, then you try to improve it with reinforcement learning, but that often makes it worse because it drifts from what it learned initially. This work shows that running both training methods in parallel, rather than sequentially, eliminates that problem. The immediate consequence is better performance on benchmarks, but the structural point is that this removes one of the known failure modes in end-to-end driving systems — the pattern where adding optimization actually degrades what you already had.