Strands Robots

Control, simulate, and train robots with natural language

Strands Docs ◆ MuJoCo ◆ NVIDIA GR00T ◆ LeRobot ◆ Robots Sim ◆ Project Board

strands-robots gives a Strands Agent hands. One Robot() call returns a MuJoCo simulation (default - no GPU, no hardware) or a real robot - same code, same natural-language control, both auto-joined to a peer-to-peer mesh.

from strands import Agent
from strands_robots import Robot

robot = Robot("so100")              # MuJoCo sim by default; mode="real" for hardware
Agent(tools=[robot])("pick up the red cube")

One agent, the whole robotics loop

Teleoperate a real arm to collect demos, fine-tune a policy on them, run it in sim and on hardware, hand work to a fleet peer, and expose it all on ROS 2 - one library, one mental model. Every line below is a distinct capability:

from strands import Agent
from strands_robots import Robot
from strands_robots.tools import train_policy

# 1. TELEOPERATE a real SO-101 with its leader arm and RECORD demos as a
#    LeRobotDataset (one prompt drives cameras + teleop + recording).
follower = Robot("so101", mode="real", port="/dev/ttyACM0",
                 cameras={"front": {"type": "opencv", "index_or_path": "/dev/video0"}})
follower.attach_teleop("so101_leader", port="/dev/ttyACM1", id="leader")
Agent(tools=[follower])(
    "start_recording(repo_id='me/pick', root='/tmp/pick', fps=30, "
    "task='pick up the cube'); teleoperate for 60s; stop_recording"
)

# 2. POST-TUNE a policy on those demos (LoRA fine-tune; GPU box).
train_policy(action="train", provider="lerobot_local",
             dataset_root="/tmp/pick", base_model="lerobot/smolvla_base",
             output_dir="/tmp/pick_ckpt", method="lora", steps=20000)

# 3. RUN the tuned checkpoint - same policy on a MuJoCo twin AND the real arm.
twin = Robot("so101")                                              # sim twin, no hardware
twin.run_policy(robot_name="so101", policy_provider="lerobot_local",
                policy_config={"pretrained_name_or_path": "/tmp/pick_ckpt"}, duration=10.0)
follower.start_task("pick up the cube", policy_provider="lerobot_local",
                    policy_port=None, duration=10.0)               # real arm, in-process

# 4. COORDINATE a fleet - tell a mesh peer to assist, in natural language.
follower.mesh.tell(follower.mesh.peers[0]["peer_id"], "hold the tray steady")

# 5. EXPOSE the running sim on ROS 2 - rviz / nav2 / any ros2 node can subscribe.
from strands_robots.simulation import Simulation
sim = Simulation(ros2_bridge=True); sim.create_world(); sim.add_robot("so101")
sim.step(100)   # publishes /so101/joint_states + camera image_raw on the ROS 2 graph

Step	Capability	Surface
1	Teleop + dataset recording	`Robot(mode="real")`, `attach_teleop`, `start_recording`
2	Policy post-tuning	`train_policy` (LeRobot / GR00T trainers)
3	Sim + hardware policy rollout	`run_policy` (sim), `start_task` (hardware)
4	Fleet coordination	`robot.mesh.tell` / `robot_mesh` tool
5	ROS 2 interop	`Simulation(ros2_bridge=True)`, `use_ros`

Steps 1 and 3-real need hardware; step 2 needs a GPU. Everything runs in sim with no hardware (Robot("so101")), so you can exercise the whole loop today.

Why strands-robots

Sim-first, safe by default. Robot("so100") spins up a MuJoCo world. You never accidentally drive real servos - mode="real" is an explicit opt-in.
50+ robots, 8 categories. Arms, humanoids, quadrupeds, hands, drones, bimanual rigs - resolved from a single registry with auto-download of assets.
Any policy. VLA models (NVIDIA GR00T, LeRobot ACT/Pi0/SmolVLA/Diffusion), plus classical motion planners, MPC, and scripted controllers behind one ABC.
Mesh networking built in. Every robot is a Zenoh peer. tell() another robot what to do; broadcast an E-STOP; bridge to AWS IoT Core for fleets.
64-action simulation tool. World building, physics, rendering, domain randomization, and LeRobotDataset recording - all agent-callable.
ROS 2 interop. Observe + command any ROS 2 graph (use_ros), act as a robot with no rclpy (use_rtps), or expose a running sim as a ROS node.
One mental model. Sim and hardware share the same policy interface, the same mesh, and the same natural-language control surface.

How it works

graph LR
    A[Natural Language<br/>'Pick up the red block'] --> B[Strands Agent]
    B --> C[Robot<br/>sim or real]
    C --> D[Policy Provider<br/>GR00T / LeRobot / planner / mock]
    D --> E[Action Chunk]
    E --> F[MuJoCo Sim<br/>or Hardware]
    F -->|observation| C

    classDef input fill:#2ea44f,stroke:#1b7735,color:#fff
    classDef agent fill:#0969da,stroke:#044289,color:#fff
    classDef policy fill:#8250df,stroke:#5a32a3,color:#fff
    classDef hardware fill:#bf8700,stroke:#875e00,color:#fff

    class A input
    class B,C agent
    class D,E policy
    class F hardware

Installation

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

uv pip install strands-robots

The base install is light (numpy, opencv-headless, Pillow). Pull in only the extras you need:

Extra	Installs	Use for
`sim-mujoco`	MuJoCo, robot_descriptions, imageio	Simulation (recommended starting point)
`sim-newton`	Newton, Warp, MuJoCo-Warp, trimesh	GPU-native simulation (NVIDIA GPU; batched envs, headless ray-traced render)
`lerobot`	LeRobot	Real hardware, local VLA inference, dataset recording
`molmoact2`	LeRobot + transformers, peft, scipy	MolmoAct2 transformers-native VLA (needs lerobot from source until PyPI >= 0.5.2)
`groot-service`	pyzmq, msgpack	NVIDIA GR00T inference client
`curobo`	(empty; install cuRobo from source)	In-process collision-aware motion planning (CUDA GPU)
`wbc`	onnxruntime	GR00T Whole-Body-Control (SONIC) humanoid locomotion - in-process ONNX, no GPU
`motionbricks`	torch + vector-quantize-pytorch, pytorch-lightning, hydra-core (install `motionbricks` from source)	NVIDIA MotionBricks generative kinematic motion for the G1 - in-process torch, composes with `wbc`
`mesh`	eclipse-zenoh, json5	Peer-to-peer robot mesh
`mesh-iot`	awsiotsdk, awscrt, boto3	AWS IoT Core mesh transport for fleets
`device-connect`	device-connect-edge, device-connect-agent-tools	Device-aware networking - discovery, RPC, events, safety (falls back to the built-in mesh if absent)
`benchmark-libero`	libero	LIBERO benchmark evaluation
`all`	everything above	Kitchen sink

# Most users start here:
uv pip install "strands-robots[sim-mujoco]"

# Real hardware + local policies:
uv pip install "strands-robots[sim-mujoco,lerobot]"

# MolmoAct2 VLA (lerobot from source until a PyPI >= 0.5.2 ships PR #3604):
uv pip install "strands-robots[molmoact2]" \
    "lerobot[feetech] @ git+https://github.com/huggingface/lerobot.git"

# Everything:
uv pip install "strands-robots[all]"

From source:

git clone https://github.com/strands-labs/robots
cd robots
uv pip install -e ".[all,dev]"

Quick starts

Simulation (no GPU, no hardware)

from strands import Agent
from strands_robots import Robot

robot = Robot("so100") # MuJoCo simulation
agent = Agent(tools=[robot])
agent("Wave the arm using the mock policy for 200 steps, then render a top-down view")

Robot("so100") returns a Simulation instance - the full 64-action simulation AgentTool. Drive it in natural language through an Agent, call its methods directly (robot.render(camera_name="topdown")), or dispatch an action by calling it (robot(action="render", camera_name="topdown")). See Simulation.

Note: Robot("so100") already creates the world and adds the robot for you. Do not call create_world() again on the returned instance - it will error with "World already exists." The create_world() / add_robot() sequence shown in Simulation (MuJoCo) is for the low-level Simulation(...) constructor, which starts empty.

Real hardware + GR00T

from strands import Agent
from strands_robots import Robot, gr00t_inference

robot = Robot(
    "so101",
    mode="real",
    cameras={
        "front": {"type": "opencv", "index_or_path": "/dev/video0", "fps": 30},
        "wrist": {"type": "opencv", "index_or_path": "/dev/video2", "fps": 30},
    },
    port="/dev/ttyACM0",
    data_config="so100_dualcam",
)

agent = Agent(tools=[robot, gr00t_inference])

# Start the GR00T inference service (Docker, Jetson/x86 GPU)
agent.tool.gr00t_inference(
    action="start",
    checkpoint_path="/data/checkpoints/model",
    port=8000,
    data_config="so100_dualcam",
)

agent("Use so101 to pick up the red block with the GR00T policy on port 8000")

Local LeRobot policy (no inference server)

from strands_robots import create_policy

# Direct HuggingFace inference - ACT, Pi0, SmolVLA, Diffusion, ...
policy = create_policy("lerobot/act_aloha_sim_transfer_cube_human")

Teleoperation (leader arms, gamepads, WASD)

Drive any real robot - or a simulation - from one or more LeRobot teleoperators. Teleoperator() mirrors the Robot() factory; attach_teleop()

teleoperate() run the control loop.

from strands_robots import Robot, Teleoperator

# Leader arm -> follower arm (both speak {motor}.pos -> zero config)
follower = Robot("so101", mode="real", port="/dev/ttyACM0")
follower.attach_teleop("so101_leader", port="/dev/ttyACM1", id="leader")
follower.teleoperate()                       # Ctrl+C or stop_teleoperate()

# Earth Rover Mini+ with WASD keys (velocity keys -> zero config)
rover = Robot("earthrover_mini_plus", mode="real", robot_ip="192.168.1.151")
rover.attach_teleop("keyboard_rover")        # W/A/S/D
rover.teleoperate(block=True, duration=30)

# Cross-vocabulary or sim teleop -> supply a map_fn(action) -> action
robot.attach_teleop("keyboard_ee", map_fn=my_ik)   # EE deltas -> joint .pos
robot.teleoperate(publish=True)              # also stream over the mesh

17 teleoperators (so100/so101/koch/omx/openarm leaders, bi_* leaders, gamepad, keyboard, keyboard_ee, keyboard_rover, phone, reachy2_teleoperator, unitree_g1, homunculus arm/glove) drive 14 robots. Zero-config when action keys match; otherwise pass map_fn. Full matrix + recipes: Teleoperation docs.

Recording & streaming datasets

The physical-AI data loop, end to end: record a LeRobotDataset from sim or hardware, stream it straight back for eval/training (no full download), and optionally dump it to a mutable Hugging Face Storage Bucket. Needs the lerobot extra (which bundles datasets + av + torchcodec).

from strands import Agent
from strands_robots import Robot

sim = Robot("so100", mesh=False)
agent = Agent(tools=[sim])

# 1. COLLECT — one natural-language prompt drives scene + cameras + policy + record.
agent(
    "Create a world with the so100 robot, add a red cube and a front camera, "
    "start recording (repo_id='local/demo', root='/tmp/demo', fps=30, "
    "overwrite=True, task='pick up the red cube'), run the mock policy for "
    "60 steps, then stop recording."
)

# 2. STREAM — read it back lazily; camera frames decode on the fly from the MP4
#    shards, state/action from parquet. Nothing is re-materialized to disk.
reader = sim.stream_dataset("local/demo", root="/tmp/demo", shuffle=False)
for frame in reader:
    frame["observation.images.front"]   # (3, H, W) tensor, decoded from video
    frame["observation.state"]          # joint vector
    frame["action"]
    break

stream_dataset() is the in-process read counterpart to start_recording/stop_recording. For full training, the upstream trainer uses the same engine — python -m lerobot.scripts.train dataset.repo_id=... dataset.streaming=true.

Verify episode integrity. A recording's ground truth is the parquet under meta/episodes/, not the count a model narrates while collecting. Collect episodes with a deterministic Python loop (one run_policy(..., n_episodes=1) plus save_episode() per episode) rather than trusting a model to count its own tool calls, then confirm the dataset holds the episodes you intended - in-process or from the shell:

sim.verify_dataset_episodes(expected=20)   # reads parquet; status="error" on a mega-episode

# exit 0 = pass, 1 = fail, so it drops straight into CI as a dataset gate
strands-robots verify-dataset /tmp/demo --expected 20

This catches the "mega-episode" corruption class - a run that buffered every frame into one episode_index=0 episode while reporting 20/20 - plus meta/info.json vs parquet drift and zero-length episodes.

Dump to a Storage Bucket during collection (mutable, Xet-deduplicated — the Phase 1/2 collection target that avoids git-LFS history bloat) with one kwarg:

sim.stop_recording(bucket="your-org/robot-fave")   # → hf://buckets/your-org/robot-fave/demo

Requires the hf CLI (pip install -U huggingface_hub + hf auth login).

Proprio-only / no video (e.g. edge devices without a torchcodec wheel): sim.stream_dataset(repo_id, drop_videos=True) streams state/action only and never touches the video decoder.

macOS note (zero-touch). torchcodec links ffmpeg via @rpath, and Homebrew's ffmpeg (/opt/homebrew/lib) is not on the default dyld search path — so video decode would normally fail with Library not loaded: @rpath/libavutil.NN.dylib. On import strands_robots we auto-detect this and put Homebrew's ffmpeg on DYLD_FALLBACK_LIBRARY_PATH (re-exec'ing the interpreter once for a plain script run; never inside Jupyter/REPL/pytest, where it just prints the one-line export to run). It's a no-op off macOS, without torchcodec, or when the var is already set. Disable with STRANDS_ROBOTS_NO_DYLD_SHIM=1. See examples/06_agent_collect_and_stream.py.

See also Recording & datasets for the DatasetRecorder direct API and append/resume workflow.

The `Robot()` factory

Robot() is a factory, not a wrapper - you get the real backend instance back with all its methods.

Robot("so100")                       # mode="sim"  (default, safe)
Robot("so100", mode="real")          # explicit hardware opt-in
Robot("so100", mode="auto")          # probe USB for servos, fall back to sim
Robot("my_arm", urdf_path="arm.xml") # bring your own MJCF/URDF

Parameter	Type	Default	Description
`name`	`str`	required	Robot name or alias (see Supported robots)
`mode`	`str`	`"sim"`	`"sim"`, `"real"`, or `"auto"` (case-insensitive)
`backend`	`str`	`"mujoco"`	Sim backend (Isaac/Newton on the roadmap)
`urdf_path`	`str`	`None`	Explicit MJCF/URDF path (skips registry lookup)
`cameras`	`dict`	`None`	Camera config (`mode="real"` only)
`position`	`list[float]`	`[0,0,0]`	Spawn position in the sim world
`data_config`	`str`	name	Observation/action schema name
`mesh`	`bool`	`True`	Auto-join the Zenoh mesh

Safety/validation rules:

Defaults to sim. Real hardware is always an explicit mode="real".
cameras= is rejected in sim mode - add sim cameras via the add_camera action after creation.
Unknown robot names raise ValueError unless you pass urdf_path=.
STRANDS_ROBOT_MODE overrides detection; a typo'd value logs a warning and falls back to sim.

Supported robots

50+ robots across 8 categories, resolved from registry/robots.json. Assets (MJCF + meshes) auto-download from robot_descriptions / MuJoCo Menagerie on first use. List them at runtime with from strands_robots import list_robots; list_robots().

Category	Count	Robots
Arm	22	so100, so101, koch, omx, panda, fr3, fr3_v2, ur5e, ur10e, xarm7, kinova_gen3, kuka_iiwa, sawyer, piper, yam, z1, vx300s, wx250s, arx_l5, openarm, hope_jr, dynamixel_2r
Humanoid	18	unitree_g1, unitree_h1, unitree_h1_2, apollo, talos, reachy2, rby1, fourier_n1, booster_t1, adam_lite, asimov_v0, cassie, elf2, jvrc, op3, open_duck_mini, toddlerbot_2xc, toddlerbot_2xm
Mobile	13	spot, go1, unitree_go2, unitree_a1, aliengo, anymal_b, anymal_c, stretch, stretch3, lekiwi, tiago_dual, earthrover, robot_soccer_kit
Hand	8	shadow_hand, shadow_dexee, allegro_hand, leap_hand, ability_hand, aero_hand, robotiq_2f85, robotiq_2f85_v4
Bimanual	3	aloha, bi_openarm, trossen_wxai
Aerial	2	crazyflie, skydio_x2
Expressive	1	reachy_mini
Mobile manip	1	google_robot

Hardware-capable (drivable with mode="real" via LeRobot): so100, so101, koch, omx, hope_jr, aloha, bi_openarm, reachy2, unitree_g1, lekiwi, earthrover. All are simulatable.

Adding a robot

There are two paths, depending on whether the robot needs project-specific metadata:

Standard robot_descriptions robot (zero config). Any MJCF robot shipped by robot_descriptions resolves automatically without a robots.json entry - the asset is discovered and downloaded on first use:
```
from strands_robots import Robot, list_discoverable

sim = Robot("iiwa14")          # discovered, not in robots.json
print(list_discoverable())     # the MJCF long tail you can load directly
```
A curated robots.json entry always wins over discovery, so overriding a discovered robot later is non-breaking.
Custom or metadata-rich robot. If the robot needs a non-default joint count, hardware port, aliases, scene tweaks, or local mesh overrides, add a curated entry. For a robot that belongs in the shipped catalog, add it to registry/robots.json and open a PR. For a machine-local robot, register it at runtime instead of editing the package:
```
from strands_robots.registry import register_robot

register_robot(name="my_arm", model_xml="my_arm.xml",
               asset_dir="~/robots/my_arm", joints=7, category="arm")
```

Tools reference

Import any of these and pass to Agent(tools=[...]). Each is a Strands AgentTool returning {"status", "content"}.

Tool	Purpose
`Robot(...)`	Universal robot - sim or hardware, natural-language + async control
`run_policy`	Multi-episode policy rollout with per-episode eval + dataset recording
`train_policy`	Post-tune (fine-tune) a policy on a recorded dataset (LeRobot / GR00T trainers, full or LoRA)
`use_lerobot`	Universal LeRobot bridge - call ANY lerobot module/class/config directly (like `use_aws` wraps boto3)
`lerobot_train`	Thin local wrapper over the `lerobot-train` CLI (the engine behind `train_policy`)
`robot_mesh`	Coordinate robots over the Zenoh mesh (`tell`, `broadcast`, E-STOP)
`use_ros`	Bridge to any ROS 2 graph - list/echo/publish topics, call services (in-process rclpy)
`use_rtps`	Join a ROS 2 graph as a DDS participant - publish/echo topics, act as a robot (pure cyclonedds, no rclpy, all ROS 2 distros)
`gr00t_inference`	Manage NVIDIA GR00T inference services (Docker lifecycle)
`lerobot_camera`	OpenCV / RealSense camera discovery, capture, record
`lerobot_calibrate`	List, view, back up, restore LeRobot calibrations
`lerobot_teleoperate`	Record demonstrations, replay episodes
`pose_tool`	Store, recall, and execute named robot poses
`serial_tool`	Low-level Feetech servo / raw serial communication
`download_assets`	Pre-fetch robot MJCF + meshes into the asset cache

Robot tool actions

Action	Parameters	Description
`execute`	`instruction`, `policy_port`, `duration`	Blocking execution until complete
`start`	`instruction`, `policy_port`, `duration`	Non-blocking async start
`status`	-	Current task status
`stop`	-	Interrupt running task (emergency stop)
In sim mode the same tool exposes the 64 Simulation actions - see Simulation (MuJoCo).

GR00T inference tool actions

Action	Parameters	Description
`start`	`checkpoint_path`, `port`, `data_config`	Start inference service
`stop`	`port`	Stop service on port
`status`	`port`	Check service status
`list`	-	List running services
`find_containers`	-	Find GR00T Docker containers
`build_image` / `download_checkpoint` / `start_container`	-	Full container lifecycle orchestration

TensorRT acceleration:

agent.tool.gr00t_inference(
    action="start",
    checkpoint_path="/data/checkpoints/model",
    port=8000,
    use_tensorrt=True,
    vit_dtype="fp8",     # ViT:  fp16 | fp8
    llm_dtype="nvfp4",   # LLM:  fp16 | nvfp4 | fp8
    dit_dtype="fp8",     # DiT:  fp16 | fp8
)

Camera / serial / pose / teleop tool actions

Camera - discover, capture, capture_batch, record, preview, test Serial - list_ports, feetech_position, feetech_ping, send, monitor Pose - store_pose, load_pose, list_poses, move_motor, incremental_move, reset_to_home Teleop - start, stop, list, replay

Policy providers

All policies implement one ABC - async get_actions(observation, instruction, **kwargs). The interface is deliberately agnostic about how actions are produced, so it fits both VLA models and classical controllers.

from strands_robots import create_policy

create_policy("mock")                                  # sinusoidal test actions
create_policy("groot", port=5555)                      # NVIDIA GR00T via ZMQ
create_policy("zmq://localhost:5555")                  # same, by URL
create_policy("lerobot/act_aloha_sim_transfer_cube")   # local HF inference

Provider	Backend	Notes
`mock`	none	Sinusoidal trajectories; `requires_images=False` (~10x faster)
`groot`	NVIDIA GR00T N1.5/N1.6/N1.7	Service mode (ZMQ to a Docker container) or local in-process (`model_path=`)
`lerobot_local`	HuggingFace	Direct ACT / Pi0 / SmolVLA / Diffusion inference, no server
`vera`	MIT VERA (DFoT/WAN planner + Jacobian IDM)	Two-stage video-to-action over a WebSocket GPU server (Docker); PushT + MimicGen, IK for eef-delta arms

classDiagram
    class Policy {
        <<abstract>>
        +get_actions(obs, instruction, **kwargs)
        +set_robot_state_keys(keys)
        +requires_images
        +reset(seed)
        +provider_name
    }
    class Gr00tPolicy
    class LerobotLocalPolicy
    class MockPolicy
    class YourPolicy
    Policy <|-- Gr00tPolicy
    Policy <|-- LerobotLocalPolicy
    Policy <|-- MockPolicy
    Policy <|-- YourPolicy

GR00T data configs (embodiment schemas)

A data_config defines the video + state keys GR00T expects for an embodiment. 27 ship in policies/groot/data_configs.json; the common ones:

Config	Cameras	Description
`so100` / `so101`	1 (`video.webcam`)	Single-arm, single camera
`so100_dualcam` / `so101_dualcam`	2 (front + wrist)	Single-arm, dual camera
`so100_4cam`	4 (front, wrist, top, side)	Single-arm, quad camera
`so101_tricam`	3 (front, wrist, side)	Single-arm, tri camera
`fourier_gr1_arms_only`	1 (ego)	Fourier GR-1 bimanual arms + hands
`unitree_g1`	1 (ego)	G1 upper body (arms + hands)
`unitree_g1_full_body` / `_locomanip`	-	G1 legs + waist + arms + hands
`bimanual_panda_gripper`	3	Dual Franka, EEF pose + gripper
`libero_panda`	2 (image + wrist)	LIBERO benchmark Panda
`oxe_droid` / `oxe_google` / `oxe_widowx`	1-2	Open X-Embodiment schemas
`agibot_*` / `galaxea_r1_pro`	3	AgiBot / Galaxea humanoids

Pick the config matching your robot's camera + state layout; pass it as data_config= to Robot(...), gr00t_inference(...), or create_policy("groot", ...).

Security: lerobot_local loads HuggingFace models with trust_remote_code=True (arbitrary code execution). You must opt in with export STRANDS_TRUST_REMOTE_CODE=1. Only load models you trust.

Cosmos 3 (NVIDIA omnimodal VLA - service mode)

nvidia/Cosmos3-Nano-Policy-DROID via a self-contained WebSocket client (cosmos3 / c3 / cosmos3://host:port); no openpi-client dep, no numpy<2 pin, so it composes with lerobot in one env.

Cosmos 3 server + client setup, embodiments, sim rollout

nvidia/Cosmos3-Nano-Policy-DROID served by the Cosmos Framework RoboLab WebSocket policy server. The policy client is self-contained - it speaks the server's msgpack+NumPy wire protocol directly via websockets + a vendored numpy packer (no openpi-client dependency, no numpy<2 pin), so it composes cleanly with lerobot for dataset recording in the same env.

1. Start the server (holds the GPU), from a Cosmos Framework checkout:

uv sync --all-extras --group=cu130-train --group=policy-server
python -m cosmos_framework.scripts.action_policy_server_robolab \
    --checkpoint-path nvidia/Cosmos3-Nano-Policy-DROID --port 8000
curl http://localhost:8000/healthz   # -> 200 when ready (~4 min cold)

2. Install the client (the cosmos3-service extra ships only msgpack

websockets - numpy-version agnostic):

uv pip install -e '.[sim-mujoco]'
uv pip install 'strands-robots[cosmos3-service]'

3. Use it (cosmos3, c3, cosmos3://host:port, or the HF model-id all resolve to Cosmos3Policy):

from strands_robots.policies import create_policy

policy = create_policy("cosmos3", embodiment="droid", port=8000)
policy.set_robot_state_keys([f"joint_{i}" for i in range(7)] + ["gripper"])
chunk = policy.get_actions_sync(observation, "pick up the cube")
# chunk == [{"joint_0": .., ..., "gripper": ..}, ...]  (one dict per timestep)

The droid embodiment (joint_pos/RoboArena) conditions on all three camera views and the server rejects a partial observation. Your observation_mapping must map a sim/robot camera onto each of observation/wrist_image_left, observation/exterior_image_1_left, and observation/exterior_image_2_left; an incomplete mapping raises an actionable client-side ValueError naming the missing keys before any request is sent (other embodiments such as umi/av/bridge need only observation/image):

policy = create_policy(
    "cosmos3", embodiment="droid", port=8000,
    observation_mapping={
        "wrist":     "observation/wrist_image_left",
        "exterior":  "observation/exterior_image_1_left",
        "exterior2": "observation/exterior_image_2_left",
    },
)

4. Roll out in MuJoCo - the droid embodiment drives a Franka/DROID-class arm, so use the franka (or panda) sim asset:

MUJOCO_GL=egl python examples/cosmos3_sim_rollout.py --record /tmp/c3.mp4

Embodiments: droid (10D, chunk 32, 15 fps), umi, av, bridge. If the server is not running, the policy raises a ConnectionError with the exact command to start it.

Non-VLA policies (motion planners, MPC, scripted)

The same interface fits cuRobo, MoveIt2, OMPL, MPC, and pure-IK / scripted trajectories - anything mapping (observation, goal) to joint targets. Non-VLA providers set requires_images = False (skip camera rendering) and read their goal from well-known **kwargs keys instead of parsing the instruction string:

Key	Type	Meaning
`target_pose`	`list[float]`	Cartesian goal `[x, y, z, qw, qx, qy, qz]` in base frame
`target_joints`	`dict[str, float]`	Joint-space goal keyed by joint name (rad / m)
`world_update`	`dict \| None`	Per-call world refresh for collision-aware planners

Providers MUST ignore unknown **kwargs rather than raising, so callers can pass shared keys across providers without coupling to a backend.

from typing import Any
from strands_robots.policies import Policy, register_policy, create_policy


class ReachPolicy(Policy):
    """Linear interpolation from current joint state to target_joints."""

    def __init__(self, steps: int = 32, **_: Any) -> None:
        self._keys: list[str] = []
        self._steps = steps

    @property
    def provider_name(self) -> str:
        return "reach"

    @property
    def requires_images(self) -> bool:
        return False  # joint-state only -- skip camera rendering

    def set_robot_state_keys(self, robot_state_keys: list[str]) -> None:
        self._keys = list(robot_state_keys)

    async def get_actions(self, observation_dict, instruction, **kwargs):
        target = kwargs.get("target_joints")
        if target is None:
            raise ValueError("ReachPolicy requires target_joints kwarg")
        state = observation_dict.get("observation.state", [0.0] * len(self._keys))
        out = []
        for s in range(1, self._steps + 1):
            alpha = s / self._steps
            out.append({k: (1 - alpha) * state[i] + alpha * target[k]
                        for i, k in enumerate(self._keys)})
        return out


register_policy("reach", lambda: ReachPolicy, aliases=["lerp"])
policy = create_policy("reach")

Reference non-VLA providers: MoveIt2, cuRobo, WBC/SONIC

Three reference implementations of the goal-kwarg contract above. Each has a runnable example + full install/deploy notes in its linked doc:

Provider	Alias	Runs	Goal kwarg	Needs	Docs
`moveit2`	`moveit`	ZMQ sidecar (ROS 2 / `moveit_py`, out-of-process)	`target_pose` / `target_joints`	`[moveit2]` extra (`pyzmq`, `msgpack`); a running sidecar	MoveIt2 docs
`curobo`	`cumotion`	in-process CUDA	`target_pose` / `target_joints` (+ `world_update`)	NVIDIA GPU; cuRobo from source (not on PyPI)	cuRobo source
`wbc`	`sonic`	in-process ONNX (CPU)	`target_velocity` `[vx, vy, omega]`	`[wbc]` extra (`onnxruntime`); a SONIC checkpoint	WBC docs
`motionbricks`	`motion_bricks`	in-process torch (CPU/CUDA)	`style` / `mode`, `target_velocity`, `target_heading`	`[motionbricks]` extra + `motionbricks` from source + git-LFS checkpoints	MotionBricks docs

from strands_robots.policies import create_policy

# Collision-aware planning (GPU, in-process); plan is cached, streamed per tick.
policy = create_policy("curobo", robot_config="franka.yml", action_horizon=16)
actions = policy.get_actions_sync(
    {"observation.state": [0.0, -0.79, 0.0, -2.36, 0.0, 1.57, 0.79]},
    "reach for the red block",                  # ignored by planners
    target_pose=[0.5, 0.0, 0.4, 1.0, 0.0, 0.0, 0.0],
)

Agents share one goal vocabulary across VLA and planner providers: Robot.start_task(..., policy_provider="curobo", target_pose=[...]) and mesh.tell(peer, "...", policy_provider="curobo", target_pose=[...]) flow the same target_pose / target_joints / world_update kwargs through.

Simulation (MuJoCo)

Robot("so100") (sim mode) returns a Simulation - a MuJoCo-backed AgentTool exposing 64 actions for world composition, physics, rendering, policy execution, and dataset recording. Build it directly when you want full control:

from strands_robots.simulation import Simulation

sim = Simulation(tool_name="sim", mesh=False)
sim.create_world()
sim.add_robot(name="arm", data_config="so100")
sim.add_object(name="cube", shape="box", position=[0.3, 0, 0.05])
sim.add_camera(name="topdown", position=[0, 0, 1.5], target=[0, 0, 0])

# Wrist camera: mount ON the gripper body so it tracks the arm like the real
# SO101/SO100 hardware cam. position/target are in the body's LOCAL frame.
# Body names are namespaced "<robot>/<body>" (e.g. "arm/gripper").
sim.add_camera(name="wrist", position=[0, -0.05, 0], target=[0, -0.15, 0],
               parent_body="arm/gripper")

sim.run_policy(robot_name="arm", policy_provider="mock", n_steps=200,
               control_frequency=50.0)

frame = sim.render(camera_name="topdown")   # {status, content:[text, image]}

The actions, grouped

World & scene: create_world, load_scene, replace_scene_mjcf, patch_scene_mjcf, reset, get_state, save_state, load_state, destroy, export_xml.
Robots: add_robot, remove_robot, list_robots, get_robot_state, list_urdfs, register_urdf, get_features.
Objects: add_object, remove_object, move_object, list_objects.
Cameras & rendering: add_camera, remove_camera, render, render_depth, render_all, start_cameras_recording, stop_cameras_recording, get_cameras_recording_status.
Physics: step, set_timestep, set_gravity, apply_force, raycast, multi_raycast, get_contacts, get_contact_forces, get_body_state, set_joint_positions, set_joint_velocities, forward_kinematics, get_jacobian, get_mass_matrix, inverse_dynamics, get_total_mass, get_energy, get_sensor_data, set_body_properties, set_geom_properties.
Policy: run_policy, start_policy, stop_policy, list_policies_running, replay_episode, eval_policy.
Randomization: randomize.
Recording (LeRobotDataset): start_recording, stop_recording, get_recording_status.
Benchmarks: list_benchmarks, register_benchmark_from_file, evaluate_benchmark.
Viewer: open_viewer, close_viewer.

Common footguns

Planes must be static. add_object(shape="plane") auto-sets is_static=True; passing is_static=False is a hard error.
Aim cameras. Pass target=[x,y,z] to look at a point; target == position errors.
Wrist cameras mount on a body. Pass parent_body="<robot>/gripper" to add_camera so the camera rides with the arm (realistic SO101/SO100 wrist cam). In that mode position/target are in the body's LOCAL frame, not world coordinates. Omit parent_body for a world-fixed camera.
MP4 vs dataset recording. start_cameras_recording writes plain MP4 ([sim-mujoco] only). start_recording writes a LeRobotDataset (parquet + MP4 + schema) and needs the [lerobot] extra.
Policy running → mutations blocked. While a policy runs, state-mutating actions error with "Cannot 'X' while a policy is running." Stop it first.
Horizon parameters. run_policy takes either duration or n_steps (both with control_frequency). fast_mode=True skips the between-step sleep for batch eval / data collection.
Name collisions. Objects, bodies, robots, and cameras share the MuJoCo name table. Multi-robot joints/actuators are namespaced {robot}/{joint}.

Self-healing: unknown parameters are rejected with "Unknown parameter X for action Y. Valid: [...]", missing required params produce "Action X requires parameter Y.", and vectors/dtypes are validated before MuJoCo sees them - so the agent learns the contract without crashing the process.

Third-party backends. create_simulation(name) discovers backends beyond the built-in mujoco/newton registry via Python entry points. A sibling package - e.g. strands-robots-sim, which ships the heavy Isaac Sim and Newton backends out-of-tree - registers its SimEngine subclasses under the strands_robots.backends group in its pyproject.toml, and they become available on pip install without patching this package:

[project.entry-points."strands_robots.backends"]
isaac = "strands_robots_sim.isaac.simulation:IsaacSimulation"
newton = "strands_robots_sim.newton.simulation:NewtonSimulation"
warp = "strands_robots_sim.newton.simulation:NewtonSimulation"

Built-in backends always take precedence over plugins of the same name, plugin discovery is lazy (it never slows cold import), and list_backends() returns the merged builtin + plugin set.

Mesh networking

Every Robot() and Simulation() is automatically a peer on a local Zenoh mesh - no setup. Peers on the same LAN discover each other via multicast scouting, sharing a single ref-counted zenoh.Session per process.

from strands_robots import Robot

a = Robot("so100")              # auto-joins the mesh
b = Robot("so100")              # second peer (another process)
print(a.mesh.peers)             # list[dict] - discovers b
print(a.mesh.peers_by_id[b.peer_id])   # dict[peer_id -> info] for O(1) lookup
info = a.mesh.get_peer(b.peer_id)      # None-safe single lookup

a.mesh.tell(b.peer_id, "pick up the cube")
a.mesh.emergency_stop()         # broadcast E-STOP, audited to disk

tell() routes to hardware and sim peers. Per-call policy kwargs (target_pose, target_joints, world_update) and constructor extras are forwarded end-to-end via policy_config, so a planner-style policy on a sim peer sees the goal payload it needs:

a.mesh.tell(
    b.peer_id,
    "reach for the red block",
    policy_provider="curobo",
    target_pose=[0.3, 0.0, 0.4, 1.0, 0.0, 0.0, 0.0],
    robot_name="arm_left",      # disambiguate in multi-robot sims
    duration=10.0,
)

Expose the mesh to an agent with the robot_mesh tool (peers, status, tell, send, broadcast, stop, emergency_stop, subscribe, watch, inbox). Disable globally with STRANDS_MESH=false or per-robot with Robot("so100", mesh=False). Install with uv pip install "strands-robots[mesh]".

For frictionless single-machine experiments, set STRANDS_MESH_LOCAL_DEV=1 - one env var that runs the mesh without mTLS/ACL on localhost. It defaults the auth mode to none and satisfies the insecure-acknowledgement second factor by itself, so you don't also need STRANDS_MESH_I_KNOW_THIS_IS_INSECURE=1. An explicit STRANDS_MESH_AUTH_MODE=mtls still wins. Never set STRANDS_MESH_LOCAL_DEV on a shared or production network.

AWS IoT Core transport (fleets)

For robots across networks, bridge the mesh to AWS IoT Core over MQTT5/mTLS, with Device Shadow mirroring, S3 camera offload, and account-wide Fleet Provisioning. Hardened with CA pinning, strict thing-name validation, deny-by-default IoT policy scoping, and a safety audit log. Install with uv pip install "strands-robots[mesh-iot]". See the Configuration matrix for the STRANDS_MESH_* knobs.

ROS 2 interoperability

strands-robots speaks ROS 2 from four complementary angles - a Strands agent can observe, command, be, and expose a ROS 2 system. Full guide: ROS 2 Integration / docs/ros2-integration.md.

A Strands agent (Claude Opus via Amazon Bedrock) given the use_ros tool drives a real ROS 2 turtlesim in a closed-loop square - reading pose, correcting heading, re-driving - over 43 in-process tool calls. Runnable: examples/ros2/use_ros/.

Surface	What it does	Backend	Needs sourced ROS 2
`use_ros`	List/echo/publish topics, call services on any ROS 2 graph	in-process `rclpy`	yes
`use_rtps`	Join a graph as a DDS peer and act as a robot (publish topics a real stack consumes)	pure `cyclonedds` (pip)	no - macOS/CI/Jetson, all distros
`RosBridgedRobot`	Drive a `cmd_vel`/odom ROS 2 base as a first-class strands `Robot`	`use_ros`	yes
`SimEngine(ros2_bridge=True)`	Publish a running MuJoCo sim's `joint_states` + camera `image_raw` so rviz/nav2/agents can subscribe	`rclpy`	yes

# Observe + command a live ROS 2 graph, in plain English:
from strands import Agent
from strands_robots.tools import use_ros
Agent(tools=[use_ros])("list the topics, drive /turtle1 forward, confirm the pose changed")

# Or expose a simulation as a ROS 2 node any tool can subscribe to:
from strands_robots.simulation import Simulation
sim = Simulation(ros2_bridge=True)
sim.create_world(); sim.add_robot("so101")
sim.step(10)   # publishes /so101/joint_states + camera image_raw on the ROS 2 domain

rclpy ships with a sourced ROS 2 distro (not on PyPI). The [ros2] extra adds only the pip-installable cyclonedds binding that use_rtps uses - so the pure-RTPS path needs no ROS install at all. Every surface degrades to a clear, structured error when its backend is unavailable; the default install never touches ROS 2.

Configuration

Environment variables

Variable	Description	Default
`STRANDS_ROBOT_MODE`	`Robot()` factory mode: `sim` / `real` / `auto`	`sim`
`STRANDS_ASSETS_DIR`	Robot model asset cache directory	`~/.strands_robots/assets/`
`STRANDS_TRUST_REMOTE_CODE`	Set `1` to allow HF `trust_remote_code` for `lerobot_local`	unset
`STRANDS_ROBOTS_NO_DYLD_SHIM`	Set `1` to disable the macOS auto-fix that puts Homebrew ffmpeg on the dyld path for torchcodec video streaming (see Recording & streaming datasets)	unset
`MUJOCO_GL`	MuJoCo GL backend (`egl`, `osmesa`, `glfw`)	auto
`GROOT_API_TOKEN`	API token for the GR00T inference service	unset
`STRANDS_MESH`	Set `false` to disable Zenoh mesh globally	`true`
`STRANDS_MESH_LOCAL_DEV`	Set `1` for a one-var localhost preset (auth `none`, no second factor needed)	unset
`STRANDS_ROS2_BRIDGE_I_KNOW_THIS_IS_INSECURE`	Second factor to expose a `Robot(ros2_transport="rtps")` inbound `joint_command` surface with no `dds_security_config` (DDS Security). Truthy: `1`/`true`/`yes`	unset

Mesh / IoT / GR00T-container env vars (advanced)

Variable	Description	Default
`STRANDS_MESH_AUTH_MODE`	Wire auth: `mtls` or `none` (`none` needs a second factor)	`mtls`
`STRANDS_MESH_I_KNOW_THIS_IS_INSECURE`	Second factor required to bring up `AUTH_MODE=none`	unset
`STRANDS_MESH_PORT`	TCP port for the local Zenoh router	`7447`
`ZENOH_CONNECT`	Comma-separated remote Zenoh endpoints to connect to	unset
`ZENOH_LISTEN`	Comma-separated endpoints for the local Zenoh listener	unset
`STRANDS_MESH_AUDIT_DIR`	Directory for the safety audit log (`mesh_audit.jsonl`)	`~/.strands_robots/`
`STRANDS_MESH_CA_PINS`	Additional SHA-256 CA pins (comma-separated 64-char hex)	unset
`STRANDS_MESH_DISABLE_CA_PIN`	Skip CA pin check on download path (break-glass)	`false`
`STRANDS_MESH_CAMERA_PRESIGN_TTL`	TTL (s) for S3 presigned camera URLs; capped at 3600	`60`
`STRANDS_MESH_ACL_FILE`	Path to a JSON5 Zenoh ACL file; unset = permissive default. See `examples/mesh_acl_example.json5` (role-scoped) and `examples/mesh_acl_strict_per_peer.json5` (per-peer). ⚠️ Required on any WAN/cloud router: mTLS gives identity, not least-privilege — without a topic-level ACL one device cert can read all fleet traffic and command any robot. See security docs.	unset
`STRANDS_MESH_POLICY_HOST_ALLOW`	Comma-separated allowlist of VLA policy-server hosts/CIDRs for inference	loopback only
`STRANDS_MESH_HITL_ACTIONS`	`robot_mesh` actions needing a human-in-the-loop interrupt: `all` / `none` / subset of `emergency_stop,broadcast,tell,send,stop,subscribe,watch`	actuation default
`STRANDS_MESH_SUBSCRIBE_ALLOW`	Extra Zenoh key-expr patterns the `robot_mesh` `subscribe` action may target, beyond the built-in low-impact set	shared classes only
`STRANDS_MESH_OVERRIDE_CODE`	Shared secret for e-stop resume HMAC proof; unset means no remote resume possible	unset
`STRANDS_MESH_INPUT_VALUE_ABS`	Absolute value clamp for teleop joint commands (radians)	`12.566` (4pi)
`STRANDS_MESH_INPUT_MAX_HZ`	Per-receiver teleop apply-rate ceiling (0 = unlimited)	`100`
`STRANDS_MESH_MAX_PEERS`	Peer registry cap; evicts oldest on overflow	`1024`
`STRANDS_MESH_RESUME_MAX_FAILS`	Failed resume attempts before cooldown engages	`5`
`STRANDS_MESH_RESUME_BACKOFF_S`	Cooldown (seconds) after exceeding resume fail threshold	`30`
`STRANDS_MESH_INPUT_AUDIT_EVERY`	Emit `input_stream_applied` audit event every N frames (0 = off)	`100`
`STRANDS_ESTOP_DEDUP_TTL_S`	E-stop fan-out Lambda dedup window (seconds)	`30`
`STRANDS_MESH_BRIDGE_TOPICS`	Comma-separated topic suffixes the Zenoh<->IoT bridge forwards (exact match). Unset = the safe default set (`presence,health,safety/event,safety/estop,safety/resume,cmd,response,broadcast`). High-volume topics (`state,pose,imu,odom,lidar`) and LAN-only topics (`camera,input,hand`) are deliberately NOT bridged	default set
`STRANDS_MESH_BRIDGE_TOPICS_PREFIX`	Comma-separated topic suffixes the bridge matches as a path prefix (so `response` matches `response/<turn-id>`). Extend this (not `STRANDS_MESH_BRIDGE_TOPICS`) when adding an RPC-shape topic with a per-turn tail	`response`
`STRANDS_GR00T_IMAGE`	Container image the `gr00t_inference` tool runs (must pass the image allowlist; agent cannot choose it)	`gr00t:latest`
`STRANDS_GR00T_IMAGE_ALLOW`	Extra image-name patterns (trailing `` = tag wildcard) added to the built-in allowlist (`gr00t:`, `nvcr.io/nvidia/isaac-gr00t:*`)	built-in only

Benchmark / diagnostic env vars (LIBERO, GR00T bisection)

Variable	Description	Default
`STRANDS_LIBERO_ACTION_LOG` / `_MAX`	Per-step OSC controller diagnostics	unset / `50`
`STRANDS_LIBERO_STATE_LOG` / `_MAX`	Per-step state values fed to GR00T	unset / `50`
`STRANDS_GROOT_WIRE_LOG` / `_MAX_CALLS`	Dump pre/post inference payloads to verify LOCAL vs SERVICE parity	unset / `10`

Asset cache

~/.strands_robots/
└── assets/           # auto-downloaded MJCF + meshes
    ├── trs_so_arm100/
    ├── franka_emika_panda/
    └── ...

Clear with rm -rf ~/.strands_robots/assets/; relocate with export STRANDS_ASSETS_DIR=/path/to/dir.

Benchmarks

strands-robots ships a LIBERO benchmark integration on the MuJoCo backend - byte-equivalent to upstream LIBERO at the model level, reaching success_rate >= 0.92 on libero-10/SCENE5. Register declarative benchmarks from file and evaluate policies via the list_benchmarks, register_benchmark_from_file, and evaluate_benchmark simulation actions. Install with uv pip install "strands-robots[benchmark-libero]".

Project structure

strands_robots/
├── __init__.py            # Lazy-loaded public API (Robot, Simulation, policies)
├── robot.py               # Robot() factory (sim/real/auto dispatch)
├── hardware_robot.py      # HardwareRobot - async LeRobot control
├── policies/
│   ├── base.py            # Policy ABC
│   ├── factory.py         # create_policy() + runtime registration
│   ├── mock.py            # MockPolicy (non-VLA reference)
│   ├── groot/             # NVIDIA GR00T (ZMQ/HTTP client + data configs)
│   └── lerobot_local/     # Direct HuggingFace inference (RTC, processors)
├── registry/              # robots.json (50+) + policies.json + loaders
├── simulation/
│   ├── base.py            # SimEngine ABC
│   ├── factory.py         # create_simulation() + backend registry
│   ├── models.py          # SimWorld / SimRobot / SimObject / SimCamera
│   └── mujoco/            # MuJoCo backend (64-action AgentTool)
├── mesh/                  # Zenoh mesh: core, sensors, input, audit, transport, iot
├── benchmarks/libero/     # LIBERO suite + BDDL parser + adapter
└── tools/                 # gr00t_inference, lerobot_*, pose, serial, robot_mesh

Development

uv pip install -e ".[all,dev]"

hatch run test          # unit tests
hatch run test-integ    # integration tests (GPU + model weights)
hatch run lint          # ruff check + format --check + mypy
hatch run format        # ruff check --fix + ruff format

Python 3.12+ required. See AGENTS.md for conventions and the accumulated code-review learnings.

Security

Found a vulnerability? Do not open a public issue. Follow the disclosure process in SECURITY.md (AWS VDP / HackerOne).

Note the trust_remote_code gate on lerobot_local (see Policy providers) and the mesh CA-pinning / thing-name validation controls in the Configuration matrix.

Contributing

Issues and PRs welcome. Track work on the Strands Labs - Robots project board; it is the source of truth for roadmap and follow-ups.

License

Apache-2.0 - see LICENSE.

Links

GitHub ◆ PyPI ◆ MuJoCo ◆ NVIDIA GR00T ◆ LeRobot ◆ Strands Docs

Name		Name	Last commit message	Last commit date
Latest commit History 503 Commits
.github/workflows		.github/workflows
docs		docs
examples		examples
strands_robots		strands_robots
tests		tests
tests_integ		tests_integ
.gitignore		.gitignore
.python-version		.python-version
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md
SECURITY.md		SECURITY.md
mkdocs.yml		mkdocs.yml
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

Strands Robots

Control, simulate, and train robots with natural language

One agent, the whole robotics loop

Why strands-robots

How it works

Installation

Quick starts

Simulation (no GPU, no hardware)

Real hardware + GR00T

Local LeRobot policy (no inference server)

Teleoperation (leader arms, gamepads, WASD)

Recording & streaming datasets

The Robot() factory

Supported robots

Adding a robot

Tools reference

Policy providers

Cosmos 3 (NVIDIA omnimodal VLA - service mode)

Non-VLA policies (motion planners, MPC, scripted)

Simulation (MuJoCo)

Mesh networking

AWS IoT Core transport (fleets)

ROS 2 interoperability

Configuration

Environment variables

Asset cache

Benchmarks

Project structure

Development

Security

Contributing

License

Links

About

Topics

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 2

Uh oh!

Contributors

Uh oh!

Languages

The `Robot()` factory