WorldOracle tests whether world models reliably predict real robot executions

Researchers from Princeton, UT Austin, Toronto, Nvidia and AgiBot launched WorldOracle, a benchmark that scores action-conditioned world models against real executions of the same actions. Arm motion is predicted well, but object positions often diverge, and scores nearly halve over full operation sequences.

• IRASim scored 78.8 on the arm vs 29.1 on the object; Cosmos 3 scored 89.4 and 45.4
• DreamDojo fell from 50.56 to 27.08 and Cosmos 3 from 53.99 to 26.60 on full sequences
• GE-Sim 2.0 led at both horizons with 73.86 and 63.33
• The platform opened to developers on October 16 for open and closed models