LoopWAM
A closed-loop world-action model that learns from deployment rollouts spanning the full quality spectrum, including both successes and failures, through mechanisms such as source-dependent noise sampling, quality guidance and quality-weighted supervision. It enables self-improvement of general-purpose robotic policies.
Timeline (last 90 days)
The LoopWAM paper proposes a closed-loop world action model that can learn from rollouts of successful, suboptimal, and failed deployments, and after three rounds of deployment training on Robot Challenge Table 30-VR, it raised the success rate from 23.67% to 43.33%, an increase of 19.7 percentage points.
LoopWAM's training data comes from two channels: the model's own deployment rollouts, and rollouts produced by other already-deployed policies.
Source-dependent noise sampling dynamically adjusts the noise level or sampling weights according to the data source, so as to keep the behavior cloning signal and the world modeling signal balanced during training.
Quality steering operates at the video/observation level, retaining full supervision for rollouts of all quality levels; quality-weighted supervision operates at the action level, reweighting the imitation loss, so that successful segments receive higher action cloning weights, and failed segments have their share in the action loss reduced while retaining their contribution in the world modeling loss.
In LoopWAM's parallel denoising process, video tokens and action tokens interact throughout, allowing the consequences learned by the model to directly influence action generation.
In closed-loop experiments on RoboChallenge Table30-V2, after three rounds of deployment-training cycles, LoopWAM's success rate increased from 23.67% to 43.33%, an improvement of 19.7 percentage points, with a final score of 56.80.
LoopWAM is a closed-loop world-action model capable of learning from deployment rollouts across the full quality spectrum, including low-quality data.
LoopWAM achieves an overall success rate of 49.5% on RoboCasa365.
LoopWAM's above results were obtained without additional embodied pretraining, and its overall performance on the three benchmarks is the best among the compared methods.
In the shirt-folding single-task post-training scenario, after two rounds of deployment-driven improvement, LoopWAM's success rate rose from 13.20% to 55.56%.
2 more timeline events
All Facts (12)
The LoopWAM paper proposes a closed-loop world action model that can learn from rollouts of successful, suboptimal, and failed deployments, and after three rounds of deployment training on Robot Challenge Table 30-VR, it raised the success rate from 23.67% to 43.33%, an increase of 19.7 percentage points.
50%UnverifiedLoopWAM's training data comes from two channels: the model's own deployment rollouts, and rollouts produced by other already-deployed policies.
50%UnverifiedSource-dependent noise sampling dynamically adjusts the noise level or sampling weights according to the data source, so as to keep the behavior cloning signal and the world modeling signal balanced during training.
50%UnverifiedQuality steering operates at the video/observation level, retaining full supervision for rollouts of all quality levels; quality-weighted supervision operates at the action level, reweighting the imitation loss, so that successful segments receive higher action cloning weights, and failed segments have their share in the action loss reduced while retaining their contribution in the world modeling loss.
50%UnverifiedIn LoopWAM's parallel denoising process, video tokens and action tokens interact throughout, allowing the consequences learned by the model to directly influence action generation.
50%UnverifiedIn closed-loop experiments on RoboChallenge Table30-V2, after three rounds of deployment-training cycles, LoopWAM's success rate increased from 23.67% to 43.33%, an improvement of 19.7 percentage points, with a final score of 56.80.
50%UnverifiedLoopWAM is a closed-loop world-action model capable of learning from deployment rollouts across the full quality spectrum, including low-quality data.
50%UnverifiedLoopWAM achieves an overall success rate of 49.5% on RoboCasa365.
50%UnverifiedLoopWAM's above results were obtained without additional embodied pretraining, and its overall performance on the three benchmarks is the best among the compared methods.
50%UnverifiedIn the shirt-folding single-task post-training scenario, after two rounds of deployment-driven improvement, LoopWAM's success rate rose from 13.20% to 55.56%.
50%UnverifiedClosed-loop improvement has high engineering complexity, requiring a stable deployment pipeline, a reliable rollout collection mechanism, and a training framework capable of handling mixed-quality data.
50%UnverifiedLoopWAM achieves an average success rate of 94.5% on RoboTwin 2.0.
50%