A new robot world model uses the signal for training its action policy to shape a compact summary of what the robot sees. On three physical bimanual tasks, the policy completed 55 of 60 trials, compared with 48 of 60 for Fast-WAM under the same pretraining and task-specific training budgets, the authors report.
A world-action model learns a robot policy while predicting how scenes change. Reconstructing the next video frame can reward visual detail that does not help a robot grasp, fold or place something. ReWAM instead uses the policy's training signal to shape the compact visual state that guides both action choices and world predictions.
ReWAM starts with frozen DINOv3-L features from a pretrained vision model, calibrates them and compresses recent observations into a Temporal Representation Bottleneck (TRB). The bottleneck is a compact summary shared by the movement policy and the future-scene model. During training, the policy's action feedback can adjust that summary, favoring information useful for choosing movements. The future-scene model still learns to predict what happens next, but its prediction errors do not change the summary; only action-policy training feedback adjusts it. ReWAM does not pretrain a generative video model, although it does use the pretrained DINOv3-L visual features.
For the physical evaluation, the team used bimanual Piper robots to collect objects, fold clothes and unpack a lunchbox. Each method received about 600 hours of self-collected embodied pretraining and 10,000 task-specific post-training steps. ReWAM succeeded in 55 of 60 attempts (91.7%); Fast-WAM succeeded in 48 of 60 (80%). Each task contributed 20 trials, so this is a small test across three task setups, not evidence of broad generalization.
In the RoboTwin 2.0 simulation benchmark, the paper reports 93.6% success for ReWAM in both clean and randomized settings, leading its strongest comparison by 0.2 and 1.4 percentage points, respectively. The study is an arXiv preprint, and its repository says code, pretrained models and documentation are still coming soon, limiting independent reproduction for now.
Sources
Image credit: Figure 1 from the ReWAM paper, used under CC BY 4.0.