WorldOracle, a robotics world-model evaluation platform, opened submissions with 335 real-operation trajectories; approximately 40% of interaction segments contain failures or suboptimal actions.

WorldOracle opened evaluation submissions for developers of action-driven world models. Researchers from AGIBOT, NVIDIA, Princeton University, the University of Texas at Austin, and the University of Toronto jointly developed the project.

The platform evaluates how accurately models predict the outcomes of robot actions. It compares generated prediction videos with actual robot executions, focusing on physical-interaction outcomes rather than primarily on video-generation quality.

Its initial collection contains 335 real-operation trajectories, including fruit grasping, towel hanging, pouring water, and opening and closing drawers. Approximately 40% of interaction segments contain failures or suboptimal actions; that percentage applies to segments, not necessarily to complete trajectories.