Building the open foundations for agents that act in the physical world.
Embodied capability is emerging in language models. On RoboDojo, Qwen3.5 to 3.7 Flash all score 0.00, and Qwen3.8 Flash reaches 31.39.
Read the research →A unified harness for robot use. uni-robot-agent connects open models, simulators, and real robots through shared, independently configurable tools. Designed for generalizable agentic training.
Robot-use agent scores on embodied leaderboards are climbing fast. 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. We propose an evaluation protocol that closes them.
OpenRUA