Clarify static apple demo reset timing#689
Conversation
There was a problem hiding this comment.
📋 Documentation Review
This PR improves the static apple demo documentation by clarifying the reset timing during teleoperation and updating the evaluation commands to use the correct policy type.
✅ Changes Look Good
Teleoperation Clarifications (step_2_teleoperation.rst):
- The added guidance about waiting for the simulation to freeze before pressing Reset is helpful. This should prevent users from inadvertently saving incomplete demonstrations.
- The new "Wait for success freeze" bullet point reinforces this important workflow step.
Evaluation Updates (step_4_evaluation.rst):
- The policy type change from
ActionChunkingClientSidePolicytoGr00tRemoteClosedloopPolicywith the added YAML config path appears to be a necessary correction. - Removing the outdated notes about
ActionChunkingClientSidePolicyvsGr00tRemoteClosedloopPolicyreduces confusion. - The timeout update (30s → 6s) and episode count adjustment (4 → 5) align the documentation with current behavior.
💭 Minor Observations
-
Timeout Change (30s → 6s): This is a significant reduction. If this reflects actual system behavior, great. If users might still need the longer timeout in certain scenarios, consider mentioning this is configurable or noting when longer timeouts might be needed.
-
Removed Troubleshooting Content: The removed
ModuleNotFoundErrortroubleshooting and action shape mismatch guidance was quite detailed. If these issues can still occur with the new policy type, users might benefit from equivalent guidance. If these issues are now obsolete with the updated approach, the removal is appropriate. -
Empty Line at L129: There appears to be a double blank line after the policy runner command block where content was removed. This is a minor RST style nit.
✔️ Summary
This is a focused documentation improvement that should help users avoid a common pitfall (resetting too early during recording) and provides accurate policy evaluation commands. The changes are clear and well-scoped.
Automated review by Isaac Lab Review Bot
Greptile SummaryThis is a documentation-only PR that clarifies the static apple demo workflow in two areas: reset timing during teleoperation data collection, and corrected policy evaluation commands.
Confidence Score: 4/5Safe to merge once the parallel-eval client-class change is confirmed against the implementation. The teleoperation clarifications are straightforward and correct. The evaluation file removes the explicit note that Gr00tRemoteClosedloopPolicy is single-env only and now uses it for --num_envs 5; if the implementation was not updated in tandem, operators following the parallel eval command will get incorrect or broken behavior with no in-doc warning. docs/pages/example_workflows/static_apple/step_4_evaluation.rst — the parallel eval client-class change warrants a quick cross-check against the current Gr00tRemoteClosedloopPolicy implementation to confirm it handles vectorized environments. Important Files Changed
Sequence DiagramsequenceDiagram
participant Op as Operator
participant Sim as Simulation
participant HDF5 as HDF5 File
Op->>Sim: Complete task (place apple on plate)
Note over Sim: Success condition triggered
Sim-->>Sim: Auto-freeze / num_success_steps recording
Sim->>HDF5: Save demonstration
Sim-->>Sim: Auto-reset simulation
Sim-->>Op: Signal: ready for next demo
Note over Op: Wait here (new guidance) before starting next demo
Reviews (3): Last reviewed commit: "fix" | Re-trigger Greptile |
There was a problem hiding this comment.
📋 Documentation Review
This PR improves the static apple demo documentation by clarifying reset timing during teleoperation and updating the evaluation commands to use the correct policy type.
✅ Changes Look Good
Teleoperation Clarifications (step_2_teleoperation.rst):
- The added guidance about waiting for the simulation to freeze before pressing Reset is clear and actionable. This should prevent users from inadvertently saving incomplete demonstrations.
- The new "Wait for success freeze" bullet point in the best practices section reinforces this important workflow step.
- The wording clarification ("may record an incomplete or failed demonstration to the dataset file") is precise.
Evaluation Updates (step_4_evaluation.rst):
- The policy type change from
ActionChunkingClientSidePolicytoGr00tRemoteClosedloopPolicywith the added--policy_config_yaml_pathflag appears to be a necessary correction for the current workflow. - Removing the outdated notes about parallel vs single-env policy types reduces confusion now that the same policy works for both.
- The timeout update (30s → 6s) and episode count adjustment (4 → 5) align the documentation with current expected behavior.
💭 Minor Observations
-
Removed Troubleshooting Content: The removed troubleshooting guidance (ModuleNotFoundError, action shape mismatch) was quite detailed. If these issues can still occur, users might benefit from equivalent guidance for the new policy type. If these scenarios are now obsolete with the updated approach, the removal is appropriate.
-
Double Blank Line (L129-130): After the policy runner command block, there is an extra blank line where the removed notes used to be. Minor RST style nit.
✔️ Summary
This is a well-scoped documentation improvement. The reset timing guidance addresses a real user pain point, and the evaluation command updates bring the docs in line with the current workflow.
Automated review by Isaac Lab Review Bot
f6abc76 to
82e713c
Compare
Summary
Doc only. Clarify static apple demo reset timing.
Detailed description