Policy

Generated from strands_robots.core.contracts.Policy at build.

Produce actions from an observation and an instruction.

A provider declares the frame it emits (action_frame) so the robot can refuse a mismatch before motion, whether it consumes images, and how many actions of each chunk the loop should execute before asking again (actions_per_step > 1 makes it chunk-emitting, and RTC engages).

Shape: ABC: subclass it, implement the abstract members.

Declared attributes

attribute	meaning
requires_images: bool = True
action_frame: UnitFrame | None = None
frame_agnostic: bool = False	True for a policy that emits in whatever frame the observation carries (the mock); preflight adopts the robot's.
control_frequency: float | None = None
actions_per_step: int = 1
provider: str = 'policy'

Members

signature	contract
set_control_frequency(hz: float) -> None	Tell the policy the executing loop's rate so latency becomes a step count.
preflight(observation_keys: Sequence[str]) -> None	Refuse, by name, an observation that lacks a key this policy needs.
async get_actions(obs: Observation, instruction: str, **kwargs: Any) -> list[Action] abstract	Return an ordered chunk of actions; a length-1 list is a single-step provider.
close() -> None	Release model handles and channels; idempotent.

Where it is used

Contracts return typed values and raise strands_robots.core.errors; the agent envelope is built in tools/. The types they exchange are in strands_robots.core.types; the frame they agree on is UnitFrame. Why the shapes are what they are: VISION.md.