Define joint and EEF action contracts for datasets

[yuecideng]

Description

This PR defines an explicit joint/EEF action contract for EmbodiChain expert and LeRobot datasets while keeping joint position as the default representation.

It preserves raw requested EEF commands through the action manager, records measured FK EEF poses as auxiliary observations, records executed controller qpos separately, and adds the Franka RLinf-facing repeated pick-and-place task variants. The joint gripper term also exposes its true 8D policy action space, including normalized gripper bounds. The contract and temporal alignment are documented in the task README and episode metadata.

This PR intentionally does not include RLinf training/runtime fixes, joint dataset regeneration or SFT checkpoints, or DexSim dependency changes. Those belong to follow-up changes after this contract is available.

Fixes #601

Type of change

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

Validation

• EmbodiChain focused tests: 55 passed.
• RLinf EmbodiChain config tests: 5 passed in the RLinf environment.
• Public API coverage: 2261/2261 exports documented.
• API documentation tests: 8 passed.
• Full-tree Black check, compileall, and git diff --check passed.

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context; no context guidance required a change.
• Public API changes are reflected in the API docs (python docs/scripts/check_api_docs.py).
• I have added tests that prove the action contract and recorder behavior.
• Dependencies have been updated, if applicable.

[greptile-apps]

Confidence Score: 5/5

[High risk] Refactors action manager and dataset recording contracts.

The PR appears safe to merge based on the reviewed changes.

Summary

The PR defines descriptor-driven policy action contracts while retaining the expert joint-action default. It also updates recording and task configurations and, since the previous review, adds pretrained-policy evaluation, rigidized-articulation support, Viewer camera tracking, and robot and solver presets. The previous Greptile threads are resolved; no new actionable finding was established.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  P[Flat policy action] --> M[ActionManager descriptors and terms]
  M --> C[Robot control]
  C --> O[Observation and rollout]
  P --> R[Policy action contract recorder]
  O --> R
  R --> D[LeRobot episode]
  E[Controller-ready expert action] --> C
  E --> X[Expert joint-action recorder]
  O --> X
  X --> D

Loading

_{Reviews (7) · Last reviewed commit: "Merge remote-tracking branch 'origin/mai..."}