Add independent physical objective evaluation

[yuecideng]

Description

Stack

• Layer: 2/2
• Base: codex/task-catalog
• Depends on: Add shared task catalog and static gallery #664
• Followed by: none

This PR now contains only the physical-objective layer originally developed with the task catalog. The catalog, CLI, static gallery, packaging foundation, and their focused tests have moved to #664.

• Add an opt-in ordered stable-region objective with per-environment progress, sustained-duration requirements, revocable final success, isolated snapshots, and partial reset.
• Keep physical results separate from program completion, segment acceptance, termination, and dataset persistence. Stable pose and velocity do not establish gripper release.
• Add explicit expert-controller trajectory replay with joint-layout and cadence validation while preserving legacy raw-policy replay.
• Add an expert/dynamic-replay evaluation runner with bounded initial XY variation, effective seeds, actual initial poses, configuration/component snapshots, and separate execution, physical, acceptance, and persistence outcomes.
• Add the repeated_pick_place objective deployment, checked-in startup evidence, public API documentation, project context, and focused tests.

Refs #106 and #655. This remains the first incremental physical-evaluation slice; learned-policy adapters, named multi-run experiments, and broader predicates remain follow-up work.

Dependencies: #664 must merge first. No new Python dependencies are introduced. Real physical qualification requires a DexSim build compatible with the repository spawn descriptor API.

Validation

• black . with Black 26.3.1: no files changed.
• python docs/scripts/check_api_docs.py: 2186/2186 exports documented.
• python .agents/skills/project-dev-context/scripts/context.py check: passed.
• python -m compileall -q embodichain embodichain_tasks: passed.
• Architecture snapshot generation and both architecture validators: passed (113 nodes, 227 edges).
• Focused objective/replay/evaluator pytest collection is blocked in the current local runtime by missing dexsim.engine; no functional assertions ran locally after the split.
• Before the history-only split, the same feature tree reported 123 passing focused tests. Fresh stacked-branch CI is authoritative.

The checked-in startup record reports an initialization error and unavailable physical outcome because the available DexSim build rejected com_quaternion. It does not claim either measured failure or success.

Type of change

• New feature (non-breaking change which adds functionality)
• Documentation update

Screenshots

Not applicable. The dependent catalog PR provides the static gallery UI.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context, updated it where needed, or explained why no update was needed.
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py).
• I have added tests that prove the feature works at the covered boundaries; local runtime limitations are documented above.
• Dependencies have been reviewed; no dependency changes are required.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Adds optional physical objective measurement to environment evaluation.

The PR appears safe to merge, with a non-blocking CUDA performance concern in milestone recording.

Findings

1.  Per-step GPU synchronization ▶

Fix with agent prompt

### Issue 1
embodichain/lab/gym/envs/objectives/ordered_placement.py:182
If the objective runs on CUDA, `torch.nonzero(reached)` forces a GPU-to-CPU synchronization on every control step, even when no milestone is reached. This can slow the simulation loop; milestone timestamps could be updated using tensor masks without extracting indices.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

The PR adds an optional ordered physical objective, expert-command dynamic replay, and a single-deployment evaluator that reports physical outcomes separately from execution and persistence.

• Recent changes add milestone timestamps, align replay action-kind metadata, and bound component snapshot reads.
• Milestone recording introduces a per-step GPU synchronization cost when objectives run on CUDA.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Control step] --> B[Physics and interval events]
  B --> C[Physical objective update]
  C --> D[Isolated progress snapshot]
  D --> E[Expert evaluation report]
  A --> F[Recorded trajectory]
  F --> G[Dynamic replay]
  G --> C

Loading

_{Reviews (5) · Last reviewed commit: "Merge origin/main into task-objective-fo..."}

[greptile-apps]

Per-step GPU synchronization

If the objective runs on CUDA, torch.nonzero(reached) forces a GPU-to-CPU synchronization on every control step, even when no milestone is reached. This can slow the simulation loop; milestone timestamps could be updated using tensor masks without extracting indices.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/lab/gym/envs/objectives/ordered_placement.py
Line: 182

Comment:
**Per-step GPU synchronization**

If the objective runs on CUDA, `torch.nonzero(reached)` forces a GPU-to-CPU synchronization on every control step, even when no milestone is reached. This can slow the simulation loop; milestone timestamps could be updated using tensor masks without extracting indices.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!