Teaching self-driving cars to learn from near-misses

Most autonomous driving systems learn only from successful expert drives, treating trajectories with similar geometry as equally safe. This misses a critical problem: two nearly identical paths can differ drastically in safety—one recoverable, one crashing. BeyondDrive learns from both successes and failures by generating hard negative examples (expert-close but dangerous trajectories) and training the model to avoid them while mimicking safe ones. The approach outperforms prior methods on standard benchmarks and transfers across different driving architectures.