Install¶

A Python environment where Robot imports, MuJoCo renders offscreen and the robot model files are on disk.

At the end of this page you have a Python environment where from strands_robots import Robot works, a MuJoCo robot renders offscreen, and the robot model files are on disk.

Requirements¶

Python >=3.12. A CPU is enough for the MuJoCo path; a GPU is only needed for the Newton and Isaac backends and for policy inference at speed.

uv venv --python 3.12 && source .venv/bin/activate
uv pip install "strands-robots[sim-mujoco]"

pip install works the same way. The bare package has four dependencies (strands-agents, numpy, opencv-python-headless, Pillow); everything else is an extra.

Pick extras¶

Most people start with sim-mujoco: MuJoCo, robot_descriptions for asset download, imageio for video, mink for inverse kinematics. Add lerobot when a physical arm arrives, mesh when a second machine does.

uv pip install "strands-robots[sim-mujoco,lerobot]"

Every extra, read from pyproject.toml at build time:

extra	installs	purpose
`cosmos3-service`	`msgpack`, `websockets`	Cosmos 3 service client
`cosmos3-diffusers`	`diffusers`, `torch`, `transformers`, `accelerate`	Cosmos 3 through diffusers
`cosmos3-sim`	`mink`, `mujoco`, `qpsolvers[quadprog]`	IK solvers for the Cosmos 3 sim path
`moveit2`	`pyzmq`, `msgpack`	MoveIt 2 planning client
`curobo`		cuRobo planning (no pins, install cuRobo yourself)
`flux3`		FLUX 3 Action policies (no pins, install flux-action yourself)
`wbc`	`onnxruntime`, `pyyaml`, `huggingface_hub`	whole body controller (ONNX checkpoints from the Hub)
`kimodo`	`torch`, `diffusers`, `transformers`, `huggingface_hub`, `accelerate`, `scipy`	Kimodo diffusion policies
`protomotions`	`onnxruntime`, `pyyaml`, `huggingface_hub`	ProtoMotions humanoid checkpoints
`holosoma`	`onnxruntime`, `huggingface_hub`	Holosoma G1 locomotion (ONNX checkpoints from the Hub)
`microduck`	`onnxruntime`, `huggingface_hub`	Microduck walking policy
`crazyflie`	`cflib`	Crazyflie radio driver
`ur`	`ur-rtde`	Universal Robots RTDE driver
`xarm`	`xarm-python-sdk`	UFactory xArm driver
`earthrover`	`requests`	Earth Rover HTTP driver
`rby1`	`rby1-sdk`	Rainbow Robotics RB-Y1 driver
`spot`	`bosdyn-client`	Boston Dynamics Spot driver
`stretch`	`hello-robot-stretch-body`	Hello Robot Stretch driver
`serial`	`pyserial`	Feetech and Dynamixel serial drivers, serial_tool, pose_tool
`lerobot`	`lerobot[feetech,dataset]`, `psutil`	lerobot drivers, teleoperation, recording, Feetech buses
`molmoact2`	`[lerobot]`, `lerobot[molmoact2]`	MolmoAct 2 policies through lerobot
`smolvla`	`[lerobot]`, `lerobot[smolvla]`	SmolVLA policies through lerobot
`groot`	`[lerobot]`, `lerobot[groot]`	GR00T N1.7 through lerobot (policy_type groot)
`sim`	`robot_descriptions`	asset download through robot_descriptions
`rl`	`[sim-mujoco]`, `torch`, `gymnasium`	reinforcement learning on the MuJoCo backend
`sim-mujoco`	`[sim]`, `mujoco`, `imageio`, `imageio-ffmpeg`, `mink`, `qpsolvers[daqp]`	MuJoCo simulation, offscreen rendering, IK (the default sim)
`sim-urdf`	`[sim-mujoco]`, `trimesh`, `pycollada`, `xacrodoc`	URDF robots compiled for MuJoCo (Collada mesh conversion)
`sim-newton`	`[sim-mujoco]`, `newton`, `warp-lang`, `mujoco`, `mujoco-warp`, `trimesh`	Newton GPU simulation on warp
`sim-mjlab`	`[sim-mujoco]`, `mjlab`, `mujoco`, `mujoco-warp`, `warp-lang`, `onnxruntime`	mjlab GPU-vectorized MuJoCo (MuJoCo-Warp), rsl_rl training, ONNX policies
`sim-isaac`	`[sim]`, `usd-core`, `imageio`, `imageio-ffmpeg`, `mujoco`, `mink`, `qpsolvers[daqp]`	Isaac Sim backend (USD assets)
`sim-gs`	`gsplat`, `plyfile`, `torch`	Gaussian splat rendering (gsplat)
`mesh`	`eclipse-zenoh`, `json5`, `cryptography`	Zenoh fleet mesh
`mesh-iot`	`[mesh]`, `awsiotsdk`, `awscrt`, `boto3`, `botocore`	AWS IoT Core bridge for the mesh
`device-connect`	`device-connect-edge`, `device-connect-agent-tools`	device-connect edge agent
`ros2`	`cyclonedds`	pure DDS (cyclonedds) ROS 2 bridge
`ollama`	`strands-agents[ollama]`	Ollama model provider for the agent
`foxglove`	`foxglove-sdk`, `mcap`	live Foxglove view and MCAP recording
`inference`	`websockets`	WebSocket inference server
`rosbridge`	`roslibpy`	rosbridge WebSocket client
`sagemaker`	`boto3`	SageMaker endpoints
`dashboard`	`fastapi`, `uvicorn[standard]`, `webauthn`, `python-multipart`, `PyJWT`, `[serial]`, `opencv-python-headless`, `huggingface_hub`	the web dashboard
`voice`	`strands-agents[bidi]`	the dashboard's speech-to-speech voice operator
`all`	`[moveit2]`, `[lerobot]`, `[sim-mujoco]`, `[sim-urdf]`, `[rl]`, `[mesh]`, `[mesh-iot]`, `[device-connect]`, `[wbc]`, `[holosoma]`, `[kimodo]`, `[protomotions]`, `[molmoact2]`, `[smolvla]`, `[groot]`, `[ollama]`, `[inference]`, `[foxglove]`, `[rby1]`, `[rosbridge]`, `[sagemaker]`, `[earthrover]`, `[xarm]`, `[spot]`, `[serial]`, `[dashboard]`	a curated bundle; GPU backends, cosmos3, ros2 and the hardware drivers stay opt-in
`dev`	`pytest`, `pytest-cov`, `ruff`, `mypy`, `pytest-timeout`, `pytest-xdist`	tests and linters

Robot models¶

Simulation loads a robot from its MJCF or URDF. Models are not in the wheel. The first Robot("so101") resolves the file through robot_descriptions and caches it under ~/.strands_robots/assets/. Set STRANDS_ASSETS_DIR to move that cache.

To fetch models ahead of time, or on a machine that will go offline, call the download tool. It is a plain function as well as an agent tool:

from strands_robots import download_assets

report = download_assets(action="status")
print("\n".join(report["content"][0]["text"].splitlines()[:3]))

You should see a count, then one line per robot with a model to fetch, [ok] when its files are present and [--] when they are not:

<available> available, <missing> missing
[ok] ability_hand         hand         PSYONIC Ability Hand (5-finger prosthetic, 11-DOF)
[ok] adam_lite            humanoid     PNDbotics Adam Lite Humanoid (26-DOF)

action="download" fetches what is missing; robots="so101,panda" narrows the set, category="arm" filters by family, force=True refetches. A download that fetched nothing returns status="error" naming the cause, not a success with zeros.

Rendering¶

MuJoCo reads MUJOCO_GL once, on first import, and the package sets it when it is unset: on Linux without a display egl or osmesa, whichever is present; on macOS cgl, the only backend, which needs a logged-in window server. Set it yourself only when that choice is wrong:

export MUJOCO_GL=egl     # Linux, headless, NVIDIA or Mesa EGL
export MUJOCO_GL=osmesa  # Linux, headless, software rendering

Check¶

strands-robots doctor

No probe touches hardware or the network; Doctor lists each and what a FAIL means. Then move a robot: First robot.

Edit page