AffordDex
AffordDex research team
AAAI 2026 research combining human motion priors with object-affordance constraints for dexterous grasping.
What you can explore
Explore where a hand should avoid contact and how the repository trains imitation, residual and vision-based policies using Shadow-hand task configurations.
Before you use it
Uses Isaac Gym and separately obtained datasets. The released task configuration is a research model, not proof of compatibility with a current commercial Shadow Hand. Results have not been reproduced by Dexhands.
Original contributors
Haoyu Zhao, Linghao Zhuang and collaborators; full author list in the original paper
License / access: No reuse license verified in the repository; check dataset and dependency terms separately
Start with the right setup.
- Environment
- Isaac Gym
- Embodiment
- Shadow configuration
What you need
Follow the Python 3.10 environment and obtain the required UniDexGrasp and OakInk2 data.
First useful check
Inspect the data-processing, grasp-generation and evaluation stages separately.
Where it stops
Dataset permissions and code permissions are separate; no code reuse license was verified.
Keep exploring
UniDexGrasp++
A generalist-and-specialist training approach for dexterous grasping across object geometries.
OakInk2
Bimanual human manipulation recordings with task structure, hand–object motion and object affordance annotations.
ManipTrans
Transfers dexterous bimanual human demonstrations to robots with residual learning.
Editorial reference · checked 2026-09-24