OpenWAM is an open world-action modeling framework that enables comparison of design choices in prediction and control by providing a common causal robot-video foundation. Built from Wan2.2-5B, it performs causal pretraining on over 10,000 hours of video (3.34 million trajectories) and integrates an action expert via Mixture-of-Transformers architecture supporting joint, sequential, and decoupled generation modes.
The framework achieves high success rates on four LIBERO suites and real-world bimanual tasks, with VTA success improving from 68.4% to 97.8% through causal adaptation. When only the video predictor adapts to new tasks, a frozen local-context inverse dynamics model trained on counterfactual data achieves 84.0% mean success across four held-out LIBERO-90 tasks.
For forward dynamics, counterfactual supervision reduces RGB prediction error by 34.5% and raises outcome identification from 21.1% to 71.3% among 16 same-state outcomes. Pretraining datasets include Open X-Embodiment (1,300,749 trajectories), Agi Bot World Beta, and others totaling 14,640 hours of source video.
Source: Hacker News · Summarized by HeadlinesBriefing