EVO-WAM: Evolving World Action Models through Video-Action Verification

EVO-WAM. Evolving World Action Models through Video-Action Verification. By Shiyang Zhou and thirteen others, from the Harbin Institute of Technology, Shenzhen, SLAI, and ten other institutions. Published on the twenty-ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Getting a robot policy to improve on a new task without collecting more expert demonstrations is a central challenge in robot learning. World action models, or WAMs, look like a possible source of supervision. They draw on broad video priors to predict future video and actions jointly. But what they generate cannot simply be trusted. A generated video may never show the task being completed. And even a video that looks successful can be paired with actions that do not match it, actions that would fail if a robot executed them.

EVO-WAM adapts a world action model to tasks it has not seen by having it learn from its own generated video-action trajectories. Crucially, candidate actions are never executed in an external environment. The model generates, a filter decides which of its own outputs are trustworthy, and it trains on those.

The framework has three parts. First, WAM training is augmented with state prediction and with an anchored multi-frame context, so that the model can roll out complete trajectories autoregressively without any execution feedback from outside. Second, reliable experience is identified in two checks. A vision-language model selects the prefixes of a rollout that complete the task, and an inverse dynamics model verifies that the video and the actions in those prefixes are consistent with each other. Third, the WAM is trained on the verified prefixes, generates new rollouts with the updated model, and the cycle repeats.

On seven unseen RoboTwin 2.0 tasks, average success with Cosmos3 rises from twenty-six point nine percent to sixty-eight percent, roughly two and a half times where it started. With DreamZero, it rises from twenty-eight point five percent to forty-six point four, about one point six times its starting rate.

Then come physical tasks. On three unseen long-horizon composite tasks in the real world, Cosmos3's average success rate climbs from twenty percent to seventy-six point seven percent. That is a gain of fifty-six point seven percentage points.

Two failures were named at the start: a video that never finishes the task, and a video whose actions do not match it. The two checks map onto them one to one, the vision-language model for the first and the inverse dynamics model for the second. With both in place, a model's own imagination becomes usable training data, and the cycle can run again and again without a single expert demonstration added.