We set out to build a foundation for training open robot-use agents. The first step was a unified harness for effective, efficient interaction across environments—without privileged information.
Robot-use agent scores on embodied leaderboards are climbing fast. How much of that is real progress? We find three loopholes that inflate them: unfair execution budgets, privileged task progress leaked to the agent during execution, and test-task solutions built into the agent before evaluation.
OpenRUA