Quickstart

Python 3.11 to 3.13:

python3.12 -m venv .venv && source .venv/bin/activate
pip install 'robouse[sim]'

On Linux without a display, MuJoCo renders offscreen with OSMesa or EGL, and Robo Use picks whichever works (MUJOCO_GL, if set, overrides the choice). robouse doctor says whether it can render:

$ robouse doctor
Offscreen rendering
  ok MuJoCo renders offscreen (MUJOCO_GL=osmesa, chosen because it was unset)
...

If it says MuJoCo cannot render offscreen, install OSMesa (or EGL, libegl1 libgl1) and run robouse doctor again:

sudo apt-get update && sudo apt-get install -y libosmesa6

The agent here is Claude Code with its default model, claude-sonnet-5. You need Claude Code on your PATH and a token from claude setup-token (with a Claude subscription) or an Anthropic API key in ANTHROPIC_API_KEY:

export CLAUDE_CODE_OAUTH_TOKEN=...

Another model or provider: Run agents and VLAs.

$ robouse tasks | grep '^arc-g'
arc-gravity tabletop
arc-gravity-vision tabletop

In arc-gravity the agent sees example grids of blocks before and after, works out the rule (blocks fall toward the front, like gravity), and then moves the real blocks with a gripper.

$ robouse run \
    --task arc-gravity \
    --harness claude-code \
    --out runs/quickstart
{
  "trial": "arc-gravity__claude-code__d10e94",
  "reward": 1.0,
  "outcome": "done",
  "steps": 731,
  "wall_s": 98.9,
  "exception": null
}

The agent observed the scene, moved the blocks with robo move-to and called robo done; the episode server then checked where the blocks were. The trial made 35 robo calls.

The trial's folder is runs/quickstart/ plus the trial name the run printed:

$ cd runs/quickstart/arc-gravity__claude-code__*/
$ cat verifier/reward.txt
1
$ cat episode/result.json
{
  "task": "arc-gravity",
  "seed": 0,
  "success": true,
  "success_ever": true,
  "success_mode": "final",
  "outcome": "done",
  "steps_used": 731,
  "max_steps": 1400,
  "requests": 35,
  "agent_text": "Applied gravity rule: blocks fall to the bottom ...",
  ...
}

• artifacts/recording.mp4: the video.
• agent/trajectory.json: what the agent said and ran. This one started with robo info, robo observe, then robo move-to 0.04 0.0 0.15 --grip -1.
• episode/trace.jsonl: every robo request and its response.

Every file in the folder: Concepts.