GoKawiilOpenWAM open-sources a framework for combining robot video and action prediction models
Researchers released OpenWAM, an open framework that lets developers compare and combine different 'world-action' approaches for robots—systems that predict future video, generate actions, or do both together. The team pretrained a shared video-action model, adapted from Wan2.2-5B, on roughly 3.34 million robot and human interaction trajectories totaling 14.64k hours of video, training it for 14 days on 32 NVIDIA B200 GPUs using video data alone before adding action learning.
GoKawiil's interpretation of the reporting above, not reported fact.
Robotics research has struggled to compare world-action model designs because each system typically uses its own backbone, dataset, and training setup, making it hard to isolate what actually improves performance. By standardizing the architecture and allowing components like video predictors and inverse/forward dynamics models to be mixed and matched, OpenWAM could make it easier for researchers to identify which design choices genuinely help robots predict and act more effectively. This kind of shared infrastructure may accelerate progress the way open benchmarks have in other areas of machine learning.
- OpenWAM is a framework for building and comparing 'world-action' models that predict robot video, actions, or both.
- Its base model was pretrained on 3.34 million trajectories (14.64k hours) of real and synthetic robot and human interaction video, using no action labels.
- The design allows independently trained modules—video predictors, inverse dynamics, and forward dynamics models—to be composed together.
Source: openwam.stanford.edu — Heng Yu, 2026-10-07
Published there as: “OpenWAM: An Open Framework for Composable World-Action Models”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.