WorldOracle benchmark finds world models often mispredict object motion in robot tasks
AIPrinceton, UT Austin, University of Toronto, NVIDIA and AgiBot introduce WorldOracle, a benchmark that scores action-conditioned world models against real robot executions. Across seven models, including DreamDojo, Cosmos 3, IRASim and GE-Sim 2.0, arm motion was predicted relatively well while object positions and movements often deviated from reality.
Source details
The Robot Report · Sponsored Content · therobotreport.comPublished · added here
Original title: WorldOracle Tests Whether World Models Can Reliably Predict Real-World Executions