Robo Use turns robot benchmarks into tasks that an agent harness, such as Claude Code or Codex, solves by driving a simulated robot from its shell with the robo command. An episode server runs the simulator, enforces the step budget, records a video and scores the final physical state.
One trial: the harness sends robo requests over a Unix socket to the episode server, which owns the simulator; the trial folder collects the agent's output, the episode record and the reward.
Every task is a BenchFlow task folder (task.md) with a reference solution and a no-op control (Suites lists their scores). Every trial is a folder in BenchFlow's trial layout: video, trajectory and reward. Published datasets of these tasks run with BenchFlow's bench eval run -d.
Tasks come from Meta-World, LIBERO, robosuite, Gymnasium-Robotics, dm_control, MyoSuite, HumanoidBench, ManiSkill3, RoboCasa, DexJoCo, MetaDrive, Robocurve's Inspect Robots and MuJoCo Menagerie robots, plus Robo Use's own suites. Results: Runs.