DexMimicGen
NVIDIA / DexMimicGen research team
Generated demonstrations and simulation environments for learning coordinated dexterous manipulation with two arms.
What you can explore
Explore demonstration expansion and imitation learning in the released simulation tasks.
Before you use it
These are simulated trajectories, not a real-robot deployment dataset. Check environment dependencies and the specific robot setup before reuse.
Original contributors
Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan and Yuke Zhu
License / access: Data: CC BY 4.0 · code: NVIDIA Source Code License
Understand the data before downloading.
Generated simulation trajectories · HDF5 demonstrations
data / demo_*
Episode group
One demonstration; episode length can vary.
- Timing and units
- Read the environment/controller metadata. Do not assume all environments share one action dimension or control rate.
- Permissions
- Data: CC BY 4.0. Code: NVIDIA Source Code License.
First inspection
- Download a chosen task from the official dataset repository.
- Inspect HDF5 groups and environment metadata locally.
- Use the original playback script for one episode before starting imitation learning.
Replaying recorded states and executing actions are different tests. Neither proves physical-hand transfer.
Editorial reference · checked 2026-09-22