OpenWAM Releases Modular World-Action Model Stack Pretrained on 6,400 Hours of Robot and Egocentric Video
Summary
Prior world-action model systems are monolithic—generative backbone, visual representation, architecture, information flow, inference, and training data are all tightly coupled—making it impossible to run controlled ablations on which design choices matter. OpenWAM is the first fully open, modular infrastructure designed to fix that.
Originally reported by paper
Read the original article →Original headline: OpenWAM Releases Modular World-Action Model Stack Pretrained on 6,400 Hours of Robot and Egocentric Video