Benchmark
RobotWorld
Test whether multimodal agents can control simulated robots across varied tasks and bodies.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
84 tasks covering manipulation, mobile manipulation, locomotion, driving, and aerial control. The paper was submitted October 7, 2026.
Methodology
Agents interpret observations, call robot tools, and revise actions from execution feedback. Task-specific control-step and auxiliary-call limits bound each episode; executable checks determine success at the budget boundary.