robotics
A Robot World Model Let Control Shape What It Remembered
ReWAM routed action loss into a compact bottleneck while its world model learned how that state evolved.
Summary
ReWAM routed action loss into a compact bottleneck while its world model learned how that state evolved.
ReWAM tests the representation shared by robot control and future-observation prediction. Feature calibration and a temporal bottleneck organize pretrained DINO features into compact states, while action-grounded shaping sends only action-loss gradients into that bottleneck. Without generative video pretraining, the authors report 93.6% success on RoboTwin 2.0; on RoboDojo they report an average score of 12.29 and 8.28% success after roughly 600 hours of embodied pretraining. Those numbers come from different evaluations and should not be read as one common scale.
Why it matters
ReWAM routed action loss into a compact bottleneck while its world model learned how that state evolved.
Limits and context
- Feature calibration and a temporal bottleneck organize pretrained DINO features into compact states, while action-grounded shaping sends only action-loss gradients into that bottleneck.
- Those numbers come from different evaluations and should not be read as one common scale.
Key claims
ReWAM routed action loss into a compact bottleneck while its world model learned how that state evolved.
Qualification: Feature calibration and a temporal bottleneck organize pretrained DINO features into compact states, while action-grounded shaping sends only action-loss gradients into that bottleneck.
Evidence: source-2026-09-30-009
Sources
- arXiv preprint 2609.38163arXiv · primary research
Corrections
No corrections have been recorded for this story.