Fix dynamic replay of position-velocity expert actions

[yuecideng]

Description

Fix dynamic replay for expert trajectories recorded with joint_command_mode="position_velocity".

Position-velocity actions are persisted as a flat [qpos, qvel] vector. Dynamic replay now reconstructs them as a ControllerAction with separate controller fields before calling env.step(). Single-environment broadcast also preserves tensor-valued per-environment lengths.

No linked issue was provided.

Dependencies: None.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves existing functionality)
• New feature (non-breaking change which adds functionality)
• Breaking change (existing functionality will not work without user modification)
• Documentation update

Validation

• black . passed with the installed Black 24.3.0 environment.
• python docs/scripts/check_api_docs.py passed (2298/2298 exports documented).
• Python compilation and git diff --check passed.
• Provider-isolated replay wrapper tests passed: 6 passed.
• Direct replay test collection is blocked in this environment by a pre-existing SciPy/scikit-learn ABI mismatch during DexSim import.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context; no context update was needed.
• Public API documentation coverage is aligned; this change does not add a public export.
• I added tests that prove the fix and cover replay modes.
• Dependencies have not been changed.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Fixes action decoding in environment replay simulation.

The PR appears safe to merge, with a non-blocking gap in end-to-end position-velocity replay coverage.

Findings

1.  Controller application remains untested ▶

Fix with agent prompt

### Issue 1
tests/gym/envs/test_replay_wrapper.py:191-194
The new position-velocity test checks only what `_FakeEnv.step()` receives; the fake environment stores the action but never applies it. Existing simulator-backed dynamic replay tests cover position-only and policy actions, so a mistake in applying either qpos or qvel during real position-velocity replay would go undetected. Please add a focused integration test when the simulator test environment is available.

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 reconstructs stored position-velocity expert actions as controller commands during dynamic replay and corrects single-environment length broadcasting.

• Adds provider-isolated tests for replay modes and action reconstruction.
• The new test does not verify application through the real controller boundary.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Recorded flat action] --> B{Expert position-velocity?}
  B -- Yes --> C[Split qpos and qvel]
  C --> D[ControllerAction]
  B -- No --> E[Flat action]
  D --> F[env.step]
  E --> F

Loading

_{Reviews (1) · Last reviewed commit: "fix(gym): replay position-velocity exper..."}

[greptile-apps]

Controller application remains untested The new position-velocity test checks only what _FakeEnv.step() receives; the fake environment stores the action but never applies it. Existing simulator-backed dynamic replay tests cover position-only and policy actions, so a mistake in applying either qpos or qvel during real position-velocity replay would go undetected. Please add a focused integration test when the simulator test environment is available.

Prompt To Fix With AI

This is a comment left during a code review.
Path: tests/gym/envs/test_replay_wrapper.py
Line: 191-194

Comment:
**Controller application remains untested** The new position-velocity test checks only what `_FakeEnv.step()` receives; the fake environment stores the action but never applies it. Existing simulator-backed dynamic replay tests cover position-only and policy actions, so a mistake in applying either qpos or qvel during real position-velocity replay would go undetected. Please add a focused integration test when the simulator test environment is available.

---

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!