NegWAM: Negative-sample-aware planning to strengthen the safety of a unified world-action model for end-to-end autonomous driving
Reinforcing End-to-End Autonomous Driving Safety via Negative-Aware Planning in a Unified World-Action Model
AI summary
NegWAM is a safety post-training framework for a unified world-action model for end-to-end autonomous driving, learning safety boundaries with near-expert risky negative trajectories as repulsive anchors. On navhard of NAVSIM v2, NegWAM reaches 40.4 EPDMS, a 4.1% relative improvement over the strongest scorer-free baseline; on navtest it reaches 93.3 EPDMS. Zero-shot closed-loop evaluation on HUGSIM verifies cross-benchmark generalization.