Add Run pi0.5 Policy objective wiring ExecutePolicyPi05Torch

[griswaldbrooks]

[written by AI]

needs: moveit_pro/#20521

Adds a lab_sim Objective, Run pi0.5 Policy (src/lab_sim/objectives/run_pi05_policy.xml), wiring the new ExecutePolicyPi05Torch behavior from moveit_pro PR #20521 (in-process libtorch pi0.5 VLA policy inference).

What it wires

• Moves to the Look at Table waypoint, then runs ExecutePolicyPi05Torch arm-only (manipulator joint group; gripper ports left unset) against lab_sim's two cameras (/scene_camera/color, /wrist_camera/color, the latter doubled to fill the checkpoint's third expected camera slot).
• Model artifact ports (prefix_pt_path, expert_pt_path, stats_path) point at placeholder paths under /models/pi05/ — no such checkpoint ships in this repo or in CI.

Known limitation: wiring demo, not runnable as-is

The pi0.5 base checkpoint's exported metadata declares state_dim: 32 (the base checkpoint has no fine-tune dataset, so the exporter falls back to PI05Config.max_state_dim's default). ExecutePolicyPi05Torch requires an exact match between the real robot state size and metadata.state_dim; lab_sim's manipulator has far fewer than 32 joints, so this objective will fail its first observation with the base checkpoint.

This is not a padding bug to fix: reading LeRobot's actual pi0.5 pipeline (lerobot/policies/pi05/processor_pi05.py's Pi05PrepareStateTokenizerProcessorStep and the shared NormalizerProcessorStep), state is discretized and embedded as text tokens in the prompt at its real (unpadded) length — max_state_dim is stored on that processor step but never read in its __call__. Only the flow-matching action expert has a hard-sized Linear projection that requires numeric padding to max_action_dim, and ExecutePolicyPi05Torch already implements that (unpadding the response chunk, per the moveit_pro PR). A properly fine-tuned checkpoint for lab_sim would report its own real joint count as state_dim in its exported metadata, and this objective would run against it unmodified — no code change needed on either side.

Runnable end-to-end once a matching-dims checkpoint is exported and its artifact paths are substituted in. Added to skip_objectives in objectives_integration_test.py so the "run all objectives" CI sweep doesn't try to execute it without real checkpoint files.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a239464c-f0a6-484b-98b6-545c5c21be1c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

MoveIt Pro Example WS - Objectives Integration Test Report

• lab_sim
  • ❌ jazzy: no report produced — see run logs
• hangar_sim
  • ❌ jazzy: no report produced — see run logs