Research
WorldBridge proposes a world action model for robot task generalization
WorldBridge pairs separate video and action experts with joint flow matching, reporting 96.23% on RoboTwin 2.0 full data and 82.8% on RoboCasa GR1 TableTop. Adding unlabeled task videos raised out-of-distribution success from 25.07% to 55.90% in the study.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
WorldBridge: Bridging Video and Action for Task Generalization in World Action Models