SimWAM: A Simple World Action Model for End-to-End Autonomous Driving
SimWAM uses future-video prediction only during training, avoiding future-frame generation at inference.
The paper frames that as the way to keep world-action modeling useful without costly test-time imagination. It co-trains a pretrained video expert with a lightweight action expert, using an isolated attention mask so action prediction does not depend on future frames. The authors report 91.5 PDMS on NAVSIM, lower latency than other WAM-based planners, and zero-shot transfer to nuScenes. Code and model weights are listed as available. HF Daily Papers' note
The paper frames that as the way to keep world-action modeling useful without costly test-time imagination. It co-trains a pretrained video expert with a lightweight action expert, using an isolated attention mask so action prediction does not depend on future frames. The authors report 91.5 PDMS on NAVSIM, lower latency than other WAM-based planners, and zero-shot transfer to nuScenes. Code and model weights are listed as available. HF Daily Papers' note
score 5