Suites

The package bundles 341 tasks in 14 suites; each is a subfolder of robouse tasks --path. The last two columns are the results of running every reference solution and the no-op control, with MuJoCo 3.3.0.

vision (top camera, as the agent sees it) — `vision` (top camera, as the agent sees it)

Stills from the recordings of one reference solution per suite; for vision, a camera image the agent receives.

Suite	From	Tasks	Needs	Oracle	No-op
`metaworld`	Meta-World MT50	50	`[metaworld]`	50/50	0/50
`gymrobotics`	Gymnasium-Robotics: Fetch, PointMaze	12	`[gymrobotics]`	12/12	0/12
`libero`	LIBERO, 5 tasks per suite	20	see below	20/20	0/20
`robosuite`	robosuite 1.5, five arm models	21	`[robosuite]`	21/21	0/21
`libero-style`	LIBERO-like, Robo Use tabletop	12	`[sim]`	12/12	0/12
`arc-style`	ARC-style rules on a table	11	`[sim]`	11/11	0/11
`robo-use-families`	Robo Use task families	8	`[sim]`	8/8	0/8
`safety`	harmful requests, benign twins	8	`[sim]`	8/8	0/8
`vision`	camera images, no coordinates	55	`[sim,metaworld]`	55/55	0/55
`hard`	low-level control, uncalibrated cameras	25	`[sim]`	25/25	0/25
`menagerie`	dual Franka Panda, ALOHA 2	15	`[sim]`, assets	15/15	0/15
`roboharm`	10 harm scenarios, 4 configurations	80	`[sim]`, assets	80/80	0/80
`drone`	Skydio X2 over a small town	10	`[sim]`, assets	10/10	0/10
`dexjoco`	DexJoCo: Allegro hand tool use	14	`[dexjoco]` + DexJoCo, see below	14/14	0/14

"Assets" means robouse fetch-assets, which downloads the MuJoCo Menagerie robot models once.

Solve rates of agents on these suites are on the Results page.

LIBERO#

LIBERO needs a virtual environment of its own, because it pins robosuite 1.4 and MuJoCo below 3.9. This recipe was run as written, with Python 3.12 on macOS:

python3.12 -m venv .venv-libero
source .venv-libero/bin/activate
pip install 'robouse[sim]>=0.1' 'mujoco==3.3.0'
pip install --no-deps \
  hf-libero==0.1.4 robosuite==1.4.0 bddl==1.0.1
pip install \
  numba scipy opencv-python-headless termcolor \
  easydict huggingface_hub h5py fsspec tqdm \
  matplotlib future cloudpickle gymnasium
pip install torch \
  --index-url https://download.pytorch.org/whl/cpu
robouse run \
  --task libero-object-0 \
  --harness oracle \
  --out runs/try

The first episode downloads LIBERO's scene assets (about 400 MB) into ~/.cache/libero/assets. In a new environment the first start can take more than two minutes while Python compiles the new packages, and robouse run may stop with TimeoutError: episode server did not start; run it again.

DexJoCo#

DexJoCo pins its own Python and MuJoCo, so it runs in a separate virtual environment that the episode server starts as a subprocess. Run once (about 300 MB):

pip install 'robouse[dexjoco]>=0.1.1'
git clone --depth 1 \
  https://github.com/brave-eai/dexjoco.git \
  ~/.cache/robouse/dexjoco
uv venv -p 3.11 ~/.cache/robouse/dexjoco-venv
VIRTUAL_ENV=~/.cache/robouse/dexjoco-venv uv pip install -e ~/.cache/robouse/dexjoco/dexjoco "numcodecs<0.16"
robouse run \
  --task dexjoco-photograph \
  --harness oracle \
  --out runs/try

The episode server looks for DexJoCo's Python at ~/.cache/robouse/dexjoco-venv/bin/python; set ROBOUSE_DEXJOCO_PYTHON to use another.