Robot Policies

Lifecycle

se3labs.interface.eval.policy.Policy defines the station-driven episode lifecycle:

Method	Called when	Return value
initialize(station, task)	Before each episode	Station_JointPose: the pose the arms move to before rollout
infer(observation)	On each observation tick	Station_ActionChunk, or None to continue the current chunk
reset()	After an episode	Nothing; clear per-episode state

Model initialization and access to task.instruction belong in initialize(). Inference frequency, preprocessing, and chunk length are controlled by the policy.

Observations and actions

station.arms and station.cameras describe the available hardware. Observation and action dictionaries use those same names. A bimanual station uses left and right for its arms; a single-arm station uses arm.

Joint targets are absolute angles in radians. Gripper positions are open fractions: 0 is closed, 1 is open. Each arm action contains arrays shaped (horizon, n_arm_dof) and (horizon, n_gripper_dof).

An action chunk's row i targets station time timestamp + i / rate. Use the observation's station-clock timestamps, not your computer's wall clock. All arms in a chunk share its horizon and rate. A new chunk replaces the remaining targets from the previous chunk; returning None keeps that chunk.

Task specification

A Task_Spec contains an operator-facing description and reset_instruction, a model-facing instruction, a max_rollout_time in seconds, and a Task_ScoreSheet. Its success question must be binary: yes means success. Additional questions can be binary, multiple choice, or exact unscored counts. The station operator supplies the answers. Sampled tasks also carry their initial layout, seed, and semantic parameters. The policy's initialize() returns the starting arm pose for each episode.

Episode completion and errors

Raising se3labs.interface.eval.policy.EndEpisode from infer() requests normal scoring, reset, and homing. Other inference exceptions are added to SessionReport.errors; that tick sends no new chunk. An initialization exception ends the session and propagates to the caller.

For a complete implementation, follow the Python API tutorial. Data types are listed in the Python Interface reference, and session reports in the Python SDK reference.

EndEpisode

class EndEpisode(Exception)

Raise from infer to finish normally: scoring, reset, then station homing.

Policy

class Policy(ABC)

Methods

initialize(station: Station_Spec, task: Task_Spec) -> Station_JointPose

Prepare for one episode of task on station and return the pose the arms start from.

infer(observation: Station_Observation) -> Station_ActionChunk | None

A new chunk, None to continue, or raise EndEpisode to score and home.

reset() -> None

Clear per-episode state. Called once after the episode ends. No-op by default.