DexLearn

Jiayi Chen / DexLearn contributors

Learning-based dexterous grasp-synthesis baselines used in BODex and Dexonomy, including diffusion and normalizing-flow approaches.

What you can explore

Explore the documented training, checkpoint sampling and evaluation through DexGraspBench. The examples include Shadow-hand grasp configurations.

Before you use it

The README does not support training the grasp-type predictor. Grasp synthesis is not a complete hardware control stack; verify model and data compatibility. Dexhands has not reproduced these results.

Original contributors

Jiayi Chen and collaborators; see the BODex and Dexonomy citations in the original repository

License / access: CC BY-NC 4.0 for the work and dataset, per the README; dependencies have separate terms

Official code, checkpoints & license statement

FROM REFERENCE TO FIRST EXPERIMENT

Start with the right setup.

INPUTObject data and grasp-training configuration

OUTPUTGenerated grasp candidates

Environment: PyTorch / DexGraspBench
Embodiment: Released Shadow configuration

What you need

Use Python 3.10, the documented dependencies, data and released checkpoints.

First useful check

Run a released evaluation configuration before retraining.

Where it stops

Grasp-type predictor training is not included. The README specifies CC BY-NC 4.0 for the work and dataset.

Original setup instructions ↗Work through the practical guide →

Documentation reviewed; not independently reproduced · reviewed 2026-09-24. Record the exact source revision when reproducing it.

Editorial reference · checked 2026-09-24

The world of dexterous hands.