In a real-robot benchmark, LoopWAM's overall task success rose from 23.67% to 43.33% over three deployment-and-training rounds, researchers report in a version 1 preprint posted Oct. 13. The model, developed by researchers from Tsinghua University, Dexmal and Loopeva, was tested on 30 tabletop tasks across four robot types.
When a robot attempts a task, a failed run still records what happened to the scene, even though its actions may not be worth copying. LoopWAM is designed to use both kinds of information: it predicts robot actions and the video that should follow them, then gives trajectories different quality scores to guide how they train the model.
For a failed run, the observed video still teaches the model about the physical outcome, but the failed actions do not directly train it to repeat those moves. Successful runs also guide action learning, while expert demonstrations receive the highest quality score. The researchers' scoring scheme uses success and a manual assessment of movement smoothness. This lets the system learn from a mistake's consequences without treating the mistake as a good action.
The real-robot test, RoboChallenge Table30-V2, spans ALOHA, DOS-W1, ARX5 and UR5 platforms. The team began with an expert-trained multitask model for each robot type. In each of three rounds, it collected about 100 deployment trajectories per task, then retrained using the expert demonstrations and accumulated robot runs. Reported overall success moved from 23.67% before those rounds to 28.00%, 37.67% and 43.33% after them.
The researchers also report a separate physical test on shirt folding: after task-specific expert training, two more deployment-and-training rounds raised success from 13.20% to 55.56%. The team has posted training and evaluation scripts and model checkpoints in a public GitHub repository. These are team-reported results in a new preprint, not an independent replication. Because each round adds data and training, the real-robot sequence does not isolate how much of the improvement came specifically from learning from lower-quality runs.
Sources
- LoopWAM preprint, version 1 (Oct. 13, 2026), especially Sections 3.2-4.4 and Tables 1-4
- public code and checkpoint repository
- Illustrative image: Larry D. Moore, "Tour4 Spoetzl Brewery Shiner Texas 2024", licensed under CC BY 4.0. The brewery robot is not part of the LoopWAM study.
