Run policies and record physical execution through RobotContext #39

Outcome

Run a supported policy on physical Geodude and record the episode through the same RobotContext, schema, ownership, safety, and dataset paths used in simulation.

Rationale

Policy interchangeability is credible only when the hardware backend does not require a parallel control stack. The initial demonstration should be deliberately bounded, should separate learned behavior from runtime safety responsibility, and should make sim-to-real differences observable rather than hiding them.

Scope

• Validate the policy, observations, calibration, action reference and bounds, rate, runtime capabilities, and hardware profile before acquiring authority.
• Support non-commanding shadow evaluation, optionally paired with a separate MuJoCo context, followed by an explicitly enabled guarded rollout.
• Run the policy through the same synchronous policy contract whether its implementation is local or supplied by a future out-of-process adapter.
• Record aligned observations, source timestamps, requested actions, applied commands, runtime events, safety states, profile and calibration fingerprints, provenance, divergence where paired simulation is used, and episode outcome.

Non-goals

• Training a new policy as part of deployment.
• Broad autonomous operation or online learning.
• Letting model output or simulator agreement bypass command validation, ownership, hardware controllers, or operator authorization.
• Requiring remote inference or a paired simulator for the first guarded rollout.

Acceptance criteria

• The same policy adapter and schema run in MuJoCo, replay, shadow, and guarded hardware modes.
• Invalid, late, stale-input, ambiguously referenced, or non-finite actions cannot reach the hardware command interface.
• Shadow mode is structurally non-commanding, and a paired simulator cannot acquire or confer hardware authority.
• Recorded physical episodes are readable through the supported LeRobot adapter and resolve to the exact description, hardware profile, calibration, policy, and runtime provenance.

Reproducible evidence

• Archive shadow and guarded-run traces, schema and freshness reports, dataset validation, and an operator run report.
• If paired simulation is used, archive the predicted-versus-observed divergence artifact with both clocks and model fingerprints.
• Provide replay fixtures for CI coverage of identical policy, deadline, stale-observation, and failure paths.

Dependencies

• Implement PolicyRunner #25
• Implement the LeRobot adapter #27
• Record execution as LeRobotDataset episodes #28
• Implement Ros2Runtime for Geodude #35
• Define Geodude hardware profiles, calibration, and semantic mappings #36
• Implement hardware fault, safety, ownership, and shadow-mode behavior #37