You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This PR adds and wires up Franka neural motion generation benchmark coverage, including planner-level trials and the Franka pick-place demo/benchmark path.
It includes:
Franka planner benchmarks for demo-matched offsets and reachable-FK waypoint banks.
Franka pick-place demo/benchmark support with selectable IK/NMG planners and automated demo execution.
Neural planner asset resolution and documentation updates for local/HuggingFace checkpoints.
Atomic-action trajectory integration updates and tests for planner, trajectory, solver, and benchmark behavior.
Motivation: make it easier to compare IK-style planning against the experimental NMG backend on Franka, while keeping downstream pick-place validation separate from planner-only metrics.
Dependencies: adds huggingface-hub for neural planner checkpoint download support.
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
Screenshots
N/A
Checklist
I have run the black . command to format the code base.
I have made corresponding changes to the documentation
I have added tests that prove my fix is effective or that my feature works
After switching the demo to the Franka arm + DH_PGI gripper, ik_interpolate can plan and physically lift the object, but the NMG planner still fails before replay. The observed failure is at the arm approach planning stage:
NeuralPlanner failed to plan arm trajectory for env 0.
PickUp failed to plan the approach trajectory.
This suggests the issue is not the gripper contact model anymore. The failure happens before physical grasp execution.
My current hypothesis is that this is mostly an out-of-distribution generalization issue for the current franka.pt NMG checkpoint. ik_interpolate solves each EEF waypoint geometrically against the live robot model/TCP, so it naturally adapts to the DH hybrid gripper and the selected grasp pose. NMG, however, is a learned waypoint policy. The hybrid DH gripper changes the effective TCP/tool geometry and produces grasp/approach poses that differ from the training distribution. Atomic pick-place also creates task-specific waypoints from grasp affordances, which are stricter than the planner-only FK/demo-offset benchmark.
Suggested improvements:
Add clearer NMG failure diagnostics: checkpoint path, start qpos, target waypoint pose, terminal pose error, active waypoint index, and whether rollout hit max_steps.
Try neural_refine as the default neural demo mode, or add IK snapping/projection for final key waypoints.
Rerank top-K grasp candidates by NMG reachability instead of selecting only the best antipodal/IK candidate.
Add fallback behavior: use NMG for free-space motion, but use IK for grasp approach/close/lift when NMG fails.
Fine-tune or retrain NMG on the Franka + DH gripper TCP distribution and atomic pick-place waypoints.
Include tool/TCP extrinsics in the NMG observation, or randomize TCP offsets during training to improve robustness to gripper changes.
Distill successful IK/TOPPRA pick-place trajectories into NMG training data.
Evaluate NMG separately on planner-only reachability and atomic pick-place, since the latter is much more sensitive to small TCP/pose errors. Uploading Kooha-2026-06-30-11-22-25.webm…
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
atomic actionatomic action related functionalitydocsImprovements or additions to documentationenhancementNew feature or requestmotion genThings related to motion generation for robotrobotModule related to robot
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
This PR adds and wires up Franka neural motion generation benchmark coverage, including planner-level trials and the Franka pick-place demo/benchmark path.
It includes:
Motivation: make it easier to compare IK-style planning against the experimental NMG backend on Franka, while keeping downstream pick-place validation separate from planner-only metrics.
Dependencies: adds
huggingface-hubfor neural planner checkpoint download support.Type of change
Screenshots
N/A
Checklist
black .command to format the code base.