OSWAM is a world action model for driving and robot arm control that generates actions in one diffusion step instead of tens of steps.
How it works: It starts diffusion from structured "imagination tokens" instead of random noise, and distills future-conditioned privileged information into the one-step generation.
Speed: The paper claims about 3x faster inference than direct-policy WAM baselines, with latency as low as 40ms on an NVIDIA RTX 5090 with compiled inference.
Results: It reports 90.3 PDMS on NAVSIM and 90.00% average success on RoboTwin 2.0, comparable to or better than multi-step baselines, plus real-world tests on PiPER manipulation tasks.
Tens of steps down to one, which is the whole pitch.
Sign in to comment.
TL;DR of the paper, generated automatically.
OSWAM is a world action model for driving and robot arm control that generates actions in one diffusion step instead of tens of steps.
Tens of steps down to one, which is the whole pitch.