DexMimicGen Cartesian-Delta Policies
Private model bundle for evaluating the basic LMS Cartesian action adapter on DexMimicGen Box and Drawer tasks.
Checkpoints
| Directory | Environment | Training output | Step |
|---|---|---|---|
box/step-20000 |
TwoArmBoxCleanup |
achieved Cartesian delta | 20,000 |
drawer/step-50000 |
TwoArmDrawerCleanup |
achieved Cartesian delta | 50,000 |
Each directory is a standalone LeRobot diffusion-policy checkpoint containing
config.json and model.safetensors.
These checkpoints output a 24-dimensional two-arm action:
0:6 left arm Cartesian delta (xyz + rotation-vector xyz)
6:12 left hand action
12:18 right arm Cartesian delta (xyz + rotation-vector xyz)
18:24 right hand action
Use the companion private repository
SChaeck/dexmimicgen-action-adapter for calibration, LMS adaptation,
AdapterBench, and finalized task-success evaluation.
Access and license
This repository is private. Access does not grant permission to redistribute DexMimicGen, simulator assets, datasets, or third-party code. The adapter and its source integration retain their original non-commercial research notices.