Self-play has revolutionized the world of gaming, and now, a groundbreaking study reveals its effectiveness in the realm of autonomous driving. Through 1.6 billion kilometers of simulated driving, a new policy has been developed that outperforms previous models and displays an incredible level of realism. This achievement was made possible by Gigaflow, a sophisticated simulator that can replicate 42 years of driving experience in just one hour on a single 8-GPU node. The resulting policy not only excels in independent benchmarks but also surpasses human drivers in real-world scenarios, demonstrating unprecedented levels of performance and robustness.
https://arxiv.org/abs/2502.03349