LEAP Hand assets & trained retargeter weights + retargeting instability with public URDF

[ligerfotis]

Could you please share the LEAP Hand assets (URDF/meshes/configs) and, if possible, the trained retargeter weights used in your experiments?

I trained and evaluated exactly as described in the README, using the provided training data. Losses converge, but the LEAP hand does not track** the video and instead produces unstable/unnatural joint behavior.

LEAP URDF used: https://github.com/leap-hand/LEAP_Hand_Sim/tree/master/assets

Steps followed as in README, loss curves converge, but:

• The hand does not follow the driving motion.
• Fingers jitter or hyperextend, especially during multi-finger gestures.
• Wrist/base shows unnatural compensations.
• Overall motion looks miscalibrated compared to the video.

What I’ve checked

• Training ran without errors; losses decreased smoothly.
• URDF loaded fine; joint limits parsed correctly.
• Units/inertias look consistent.
• Tried clamping predictions to joint limits and adding smoothing → instability persists.

My hypothesis is that the public LEAP URDF may not match the one used in your experiments (frames, joint limits, actuator modeling, inertias) or the model may require normalization constants, calibration settings, or different hyperparameters not mentioned in the README.

Requests

1. LEAP assets: the canonical URDF/meshes and joint/limit definitions used in your work.
2. Pretrained weights: LEAP retargeter checkpoint + config (loss weights, architecture, inference filters).

Thanks in advance for considering a release of the LEAP assets and weights! It would make reproducing results and resolving this mismatch much easier.

[zhaohengyin]

Thanks for raising this issue. We do notice that certain LEAP hand public URDF has a much larger default joint moving limit (especially thumb) which is unnecessarily overly large that can lead to fingertip configuration space mismatch. You will need to tune that based on your application preference.

I will post an updated asset shortly.

[akuramshin]

Thanks for raising this issue. We do notice that certain LEAP hand public URDF has a much larger default joint moving limit (especially thumb) which is unnecessarily overly large that can lead to fingertip configuration space mismatch. You will need to tune that based on your application preference.

I will post an updated asset shortly.

I have attempted to restrict the joint limit range for all the fingers and especially the thumb, but the resulting retargeting is still unstable.

robot.urdf.txt