Mind2Act World Benchmark Finds 12% Success on Combined Robotics Tasks
The benchmark spans 180 simulated configurations. Its strongest agent scored higher when reasoning and execution were tested separately.

Mind2Act World tests whether robotic systems can combine reasoning with physical action, and its initial results show a gap between those abilities when tasks require both in a feedback loop.
The strongest agent in initial evaluations succeeded on 12.0% of combined Mind2Act tasks, compared with 52.7% on reasoning tasks and 39.0% on execution tasks.
The benchmark covers 30 simulated manipulation tasks drawn from five real-world application domains. Three difficulty levels and two scene settings for each task bring the total to 180 configurations. The settings are labeled “Clean” and “Random.”
Mind tasks test reasoning from task history, while Act tasks measure physical execution. Mind2Act tasks combine the two in a closed loop, where action outcomes inform the agent’s next decisions.
A diagnostic harness groups failures by reasoning, execution or the interaction between them. The benchmark’s GitHub project includes physics-verified task scoring, clean training and validation layouts, evaluation tools and leaderboards.
The project team plans to release a smaller version called Mind2Act Mini World. A new agent leaderboard and a standardized evaluation interface are also in development.