Policy providers

strands_robots ships several policy providers. The registry is the ground truth - list every provider that create_policy("<name>") accepts with:

python -c 'from strands_robots.policies import list_providers; print(list_providers())'
# ['cosmos3', 'curobo', 'groot', 'lerobot_async', 'lerobot_local', 'mock', 'motionbricks', 'moveit2', 'remote', 'vera', 'wbc', 'wbc_gait']

from strands_robots.policies import create_policy, list_policy_types, list_providers

print(list_providers())     # sorted provider names (registry ground truth)
print(list_policy_types())  # lerobot_local policy_type strings: ['act', 'diffusion', 'smolvla', ...]

policy = create_policy("mock")                                                     # always works, no model
policy = create_policy("groot", port=5555, data_config="so100_dualcam")
policy = create_policy("lerobot_local", pretrained_name_or_path="lerobot/pi0_so100")
policy = create_policy("cosmos3", embodiment="droid", port=8000)
policy = create_policy("remote", endpoint="ws://gpu-box:8765")

Providers

Every row below is a registered provider (create_policy("<name>")). The table is kept in sync with list_providers() by a regression test (tests/test_docs_policy_coverage.py), so it can never silently drift.

Provider	Class	Install extra	When to use
mock	MockPolicy	(core)	Tests, smoke checks; sinusoidal joints, no GPU. Reference minimal Policy (documented inline + custom-policies)
groot	Gr00tPolicy	groot-service	NVIDIA GR00T N1.5/N1.6/N1.7 over ZMQ
lerobot_local	LerobotLocalPolicy	lerobot	HF LeRobot in-process (ACT, Pi0, SmolVLA, MolmoAct2, ...)
lerobot_async	LerobotAsyncPolicy	lerobot-async	Offload a LeRobot policy to a GPU box over lerobot's native async-inference gRPC transport; the robot host stays light. Edge-device inference
cosmos3	Cosmos3Policy	cosmos3-service	NVIDIA Cosmos 3 omnimodal VLA over WebSocket
vera	VeraPolicy	vera	MIT VERA video-to-action (DFoT/WAN planner + Jacobian IDM) over a containerized GPU server
remote	RemotePolicy	inference	Offload a large policy to a GPU box: forward observations to a remote PolicyServer over WebSocket, get back action chunks. Edge-device inference
curobo	CuroboPolicy	curobo	NVIDIA cuRobo collision-aware motion planning, in-process CUDA (non-VLA)
moveit2	MoveIt2Policy	moveit2	MoveIt2 motion planning over a ROS 2 sidecar (ZMQ), no in-venv ROS 2 deps (non-VLA)
wbc	WBCPolicy	wbc	NVIDIA GR00T Whole-Body-Control (SONIC) Unitree G1 humanoid locomotion, in-process ONNX, no GPU (non-VLA)
wbc_gait	WBCGaitPolicy	wbc	WBC gait-clock variant: single ONNX policy, 95-dim obs + bipedal phase clock (non-VLA)
motionbricks	MotionBricksPolicy	motionbricks	Generative kinematic Unitree G1 motion (style-driven: walk/stealth_walk/...), in-process torch (non-VLA)

Policy ABC

from strands_robots.policies import Policy   # strands_robots/policies/base.py

class MyPolicy(Policy):
    # three abstract methods - must implement all:
    async def get_actions(self, observation_dict: dict, instruction: str, **kw) -> list[dict]: ...
    def set_robot_state_keys(self, keys: list[str]) -> None: ...
    @property
    def provider_name(self) -> str: ...

    # optional overrides:
    @property
    def requires_images(self) -> bool: return True   # False for state-only policies
    def reset(self, seed=None): pass                  # clear episode state; default no-op
    # sync helper provided by base: get_actions_sync(obs, instruction, **kw) -> list[dict]

Factory

from strands_robots.policies import register_policy

register_policy("my_prov", lambda: MyPolicyClass, aliases=["mp"])
policy = create_policy("my_prov")

Smart URI strings also resolve: "zmq://localhost:5555" → groot; "cosmos3://host:8000" → cosmos3.

In simulation

# Provider name + kwargs in policy_config={}
sim.run_policy(robot_name="so100", instruction="pick up the cube",
               policy_provider="groot",
               policy_config={"port": 5555, "data_config": "so100_dualcam"},
               duration=10.0)

# Pre-built instance via policy_object=
sim.run_policy(robot_name="so100", instruction="pick up the cube",
               policy_object=create_policy("groot", port=5555, data_config="so100_dualcam"),
               duration=10.0)

policy_config must be a dict (it is forwarded to create_policy with **); the same holds for the per-call policy_kwargs. Any other shape - a "port=5555" string, a list of pairs, an unparsed JSON blob - is rejected by run_policy / start_policy / eval_policy / evaluate_benchmark with a structured error naming the parameter, before any policy is created.

LerobotLocalPolicy requires export STRANDS_TRUST_REMOTE_CODE=1 (raises UntrustedRemoteCodeError otherwise).

See also

• GR00T - ZMQ server, 27 embodiments, container lifecycle.
• LeRobot Local - in-process HF models, RTC.
• LeRobot Async - offload a LeRobot policy to a gRPC PolicyServer (edge offload).
• MolmoAct2 (SO-100/101) - action/observation contract for the SO-arm checkpoints.
• Persistent worker - load once, reuse across rollouts; cache controls + telemetry.
• Cosmos 3 - NVIDIA Cosmos 3 omnimodal VLA.
• VERA - MIT video-to-action planner + Jacobian IDM over a GPU server.
• Remote - forward observations to a remote PolicyServer over WebSocket (edge offload).
• cuRobo - in-process collision-aware motion planning (non-VLA, GPU).
• MoveIt2 - ROS 2 sidecar collision-aware planning (non-VLA, no in-venv ROS 2).
• WBC - GR00T Whole-Body-Control (SONIC) G1 locomotion (non-VLA, in-process ONNX).
• WBC gait-clock variant - single-ONNX gait-clock G1 controller (non-VLA).
• MotionBricks - generative kinematic G1 motion (non-VLA, in-process torch).
• Custom policies - implement the ABC.