EVO-WAM: Self-Improving Robot Policies via Video-Action Verification

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

The authors report a framework called EVO-WAM designed to adapt robot policies to unseen tasks without requiring new expert demonstrations. The approach leverages world action models (WAMs), which jointly predict future videos and actions, to generate candidate trajectories. To address the issue that generated videos may not depict task completion or may be paired with inconsistent actions, the framework introduces a verification pipeline. This pipeline uses a vision-language model to select task-completing prefixes and an inverse dynamics model to check video-action consistency. The WAM is then iteratively trained on these verified prefixes to generate improved rollouts.

In their evaluation, the authors report that on seven unseen RoboTwin 2.0 tasks, EVO-WAM increases average success rates from 26.9% to 68.0% for the Cosmos3 model and from 28.5% to 46.4% for the DreamZero model. These gains represent approximately 2.5x and 1.6x their initial success rates, respectively. For real-world testing, the authors state that on three unseen long-horizon composite tasks, the method improves Cosmos3's average success rate from 20.0% to 76.7%, a gain of 56.7 percentage points.

Why it matters

A central challenge in robot learning is improving policies on new tasks without the high cost of collecting additional expert demonstrations. Traditional approaches often rely on supervised learning from human data or reinforcement learning in external environments, both of which can be resource-intensive. EVO-WAM offers a different path by using the model's own generated data as a source of supervision. By filtering out inconsistent or incomplete trajectories before training, the framework aims to provide a more efficient way to adapt existing world action models to new domains without external execution feedback.

Robot's take

The core strength of EVO-WAM is its ability to self-improve by verifying its own outputs, which could significantly reduce the data collection bottleneck in robotics. The reported improvements, particularly the 56.7 percentage point gain in real-world tasks, suggest that video-action consistency is a critical factor in execution success. However, the method relies heavily on the quality of the initial world action model and the accuracy of the verification models. It is not yet clear how well this approach generalizes to tasks with highly dynamic or unpredictable environments, where video generation may struggle to maintain physical plausibility. Further validation on a broader range of real-world scenarios would be needed to confirm its robustness.