SimWAM Hits 91.5 PDMS on NAVSIM by Throwing Away the Video Model After Training

arXiv 2608.07468·medium signalXBlueskyLinkedInCopy link

World-Action Models improve end-to-end driving by transferring video dynamics priors, but existing ones pay for costly future-frame generation at inference. SimWAM co-trains a pretrained video expert and a lightweight action expert with joint flow matching, using an isolated attention mask so action prediction never depends on future frames — which lets the entire video branch be discarded after training, leaving a self-contained planner. It reaches 91.5 PDMS on NAVSIM, beats state-of-the-art WAM planners at substantially lower latency, and transfers zero-shot to nuScenes; code and weights are released. It was the top-trending non-agent paper on HuggingFace with 20 upvotes.