Installation¶

Requires Python >= 3.12. Examples use uv (curl -LsSf https://astral.sh/uv/install.sh | sh); plain pip install works too.

Extras matrix¶

Extra	Pulls in	When you need it
(none)	core only - Robot factory, registry, lazy imports	Inspect the catalog, write tools
`[sim]`	`robot_descriptions>=1.23.0,<2.0.0`	Sim asset resolution without MuJoCo
`[sim-mujoco]`	`sim` + `mujoco`, `imageio`, `imageio-ffmpeg`	Any `Robot()` with default `mode="sim"`
`[lerobot]`	`lerobot>=0.6.1,<0.7.0`, `psutil>=6.0.0,<8.0.0`	`LerobotLocalPolicy` + dataset recording + the `lerobot_train` / `lerobot_teleoperate` session tools
`[groot-service]`	`pyzmq`, `msgpack`	`Gr00tPolicy` (ZMQ to a GR00T container)
`[cosmos3-service]`	`msgpack`, `websockets>=17.0`	`Cosmos3Policy` (WebSocket to Cosmos 3 server)
`[earthrover]`	`requests>=2.28.0,<3.0.0`	`Robot("earthrover", mode="real", driver="strands")` - HTTP to the earth-rovers-sdk
`[ur]`	`ur-rtde>=1.6.0,<2.0.0`	`Robot("ur5e", mode="real", driver="strands")` - RTDE to a UR controller
`[mesh]`	`eclipse-zenoh>=1.6.1,<2.0.0`, `json5`	Multi-robot mesh discovery + RPC
`[mesh-iot]`	`mesh` + `awsiotsdk`, `awscrt`, `boto3`	AWS IoT Core transport for mesh
`[all]`	20 of the 32 extras - not a union. `[cosmos3-diffusers]`, `[cosmos3-service]`, `[cosmos3-sim]`, `[crazyflie]` (GPLv3), `[curobo]`, `[microduck]`, `[ros2]`, `[sim-gs]`, `[sim-isaac]`, `[sim-newton]` and `[ur]` (compiled binding) stay opt-in	Demos, CI, exploration
`[dev]`	`pytest`, `pytest-cov`, `ruff`, `mypy`, `pytest-timeout`	Contributing

uv pip install "strands-robots[sim-mujoco]"                  # sim only
uv pip install "strands-robots[all]"                         # the 21-extra bundle
uv pip install "strands-robots[sim-mujoco,cosmos3-service]"  # Cosmos 3
uv pip install "strands-robots[sim-mujoco,lerobot,mesh]"     # pick and choose

The [sim] floor is set by the robot catalog rather than by an API: each entry in the built-in registry names the robot_descriptions submodule that fetches its MJCF and meshes, and that package gains one module per newly packaged robot. robot_descriptions 1.23.0 is the oldest release providing a module for every registered robot - on an older one, robots such as so100 and so101 have no module to import and no fallback, so Robot("so101", mode="sim") cannot resolve a model file.

The websockets floor in [inference] and [cosmos3-service] is set the same way, by what the code needs rather than by preference - and by a behaviour rather than a name. PolicyServer.stop() is documented to stop the server serving, not merely listening; through websockets 16.x Server.shutdown() closed the listening socket alone, so a client that was already connected went on being answered with action chunks after the caller was told the server stopped, which on a robot is the policy still driving the arm. websockets 17.0 closes the connections it accepted and waits for their handlers. Measured against the released wheels on unchanged sources: 16.1.1 still serves that client, 17.0 does not - so both extras declare >=17.0. (13.0 remains the floor of the API names reached for, websockets.sync.server.Server; the higher of the two wins.) They declare the same floor on purpose: an environment resolves one websockets, so two different floors would leave the lower one describing an install nobody gets.

Platform notes¶

macOS: works out of the box (arm64 + Intel).

Linux (headless / real hardware):

sudo apt install libosmesa6-dev ffmpeg
sudo usermod -aG dialout $USER   # USB serial access; re-login after

Windows: WSL2 + Ubuntu 22.04 (native Windows works for sim, not actively tested).

Jetson / aarch64 (JetPack):

uv pip install "numpy<2" "pandas==2.1.4"
uv pip install "strands-robots[sim-mujoco,lerobot]"

lerobot 0.6 pulls torchcodec on aarch64 itself (its dependency marker now covers linux aarch64 and pins the torch-ABI-matched torchcodec 0.11), so the video decoder resolves without a strands override. If torch CUDA is needed on Jetson, ensure you install from NVIDIA's index or set UV_TORCH_BACKEND=auto:

export UV_TORCH_BACKEND=auto   # resolves +cu130 wheels for Thor/Jetson
uv pip install "strands-robots[sim-mujoco,lerobot]"

MolmoAct2 on Jetson¶

MolmoAct2 checkpoints (e.g. allenai/MolmoAct2-SO100_101) resolve straight from PyPI now that lerobot >= 0.6 ships MolmoAct2Policy (it was added after lerobot 0.5.1). See LeRobot Local: MolmoAct2 for full instructions. Quick path:

# The [molmoact2] extra layers transformers, peft, scipy on top of lerobot >= 0.6;
# lerobot 0.6 pulls the aarch64 torchcodec decoder itself:
uv pip install "strands-robots[molmoact2]"

Headless rendering¶

export MUJOCO_GL=osmesa     # software rendering - Linux
export MUJOCO_GL=egl        # hardware EGL

Or in Python before first import:

import os
os.environ["MUJOCO_GL"] = "osmesa"
from strands_robots import Robot

Verify¶

doctor checks this machine the way the runtime will read it - the interpreter and package, each extra, the GL backend, the torch/torchcodec pair, the GPU, the serial and Hub credentials, and the device-connect and mesh postures - and exits non-zero if any row fails:

python -m strands_robots doctor           # run every check
python -m strands_robots doctor --list    # print the check names, probe nothing

from strands_robots import Robot

sim = Robot("so100")
sim.step()
obs = sim.get_observation("so100")
# obs is a flat dict mixing per-joint state floats and per-camera ndarrays:
#   {'shoulder_pan.pos': 0.0, ..., 'gripper.pos': 0.0, 'default': <HxWx3 uint8>}
print(list(obs.keys()))

Assets cache under ~/.strands_robots/assets/.

Environment variables¶

Env var	What	Default
`STRANDS_ASSETS_DIR`	Robot model asset cache	`~/.strands_robots/assets/`
`STRANDS_MESH_AUDIT_DIR`	Safety audit log	`~/.strands_robots/`
`MUJOCO_GL`	GL backend	auto
`STRANDS_TRUST_REMOTE_CODE`	Allow HF `trust_remote_code=True`	`false`
`STRANDS_ROBOT_MODE`	Default `Robot()` mode	`sim`
`STRANDS_MESH`	Set to `true` to opt a bare `Robot()` into the mesh; `false` disables it globally	unset (mesh off)
`GROOT_API_TOKEN`	GR00T service API token (falls back from `api_token=` kwarg)	unset