Scaling expert demonstrations alone does not help generalist robot policies improve once they are deployed, because real-world rollouts contain a mix of successes, suboptimal actions, and failures. The authors of LoopWAM argue that naively imitating every recorded action reinforces poor behavior, while discarding failures loses evidence of physical interactions and their consequences. Their method instead learns directly from this full quality spectrum, using rollouts from both the policy itself and other deployed policies.
LoopWAM is a closed-loop world-action model that unifies behavior learning and action-conditioned world modeling through source-dependent noise sampling. Quality steering and quality-weighted supervision keep full video supervision across all quality levels while regulating action imitation. Video and action tokens interact throughout parallel denoising, so learned consequences directly inform action generation.
On the RoboChallenge Table30-V2 benchmark, three deployment-and-training rounds raised success from 23.67% to 43.33%, a 19.7-percentage-point gain, with a final score of 56.80. LoopWAM also achieved 94.5% average success on RoboTwin 2.0 and 49.5% overall success on RoboCasa365 without additional embodied pretraining, outperforming compared methods across all three benchmarks. In a challenging shirt-folding task, two rounds of deployment-driven improvement after single-task post-training raised success from 13.20% to 55.56%.
These results demonstrate that closed-loop improvement from mixed-quality experience is feasible in both multi-task learning and single-task post-training settings. The paper is a single source, and no conflicting claims are reported within it.