Stanford researchers release open framework for robots that predict before they act
OpenWAM lets labs compare designs for AI models that pair video prediction with robot control.
Stanford University researchers have released OpenWAM, an open framework for building AI models that predict what a robot will see and decide how it should move, according to a paper posted Oct. 6 on arXiv. The paper is a preprint and has not been peer reviewed.
These "world-action models" pair video prediction with robot control. The researchers said existing systems change many parts at once, making it hard to tell which design choices matter. OpenWAM gives labs a shared base: a video model trained on more than 10,000 hours of robot video, which can generate video and actions together or in either order.
In simulated tests, the team reported that training on robot video raised success on one long-task benchmark from 68.4% to 97.8%. The project page shows a robot making toast and finishing a small Rubik's cube. The 10 authors, all listed with Stanford, include Li Fei-Fei and Jiajun Wu. The code is public.
Source: Yu et al., "OpenWAM: An Open Framework for Composable World-Action Models," arXiv, Oct. 6, 2026.