DexMimicGen

NVIDIA's automatic data-generation system that expands a few dozen demonstrations into tens of thousands of bimanual dexterous-hand demonstrations.

DexMimicGen is an automated data-generation system from NVIDIA Research, together with the University of Texas at Austin and UC San Diego, published at ICRA 2025; it extends MimicGen to bimanual dexterous manipulation, such as the upper body of a humanoid robot. MimicGen's approach is to split a small number of human demonstrations into per-object subtask segments, transform those segments to match new object positions, replay them in simulation, and keep only the trajectories that succeed. In two-arm tasks, the two hands sometimes act independently, sometimes need to move in sync, and sometimes must act in a specific order, so DexMimicGen adds per-arm subtask segmentation along with synchronization and ordering constraints. The paper generates about 21,000 demonstrations from 60 human demonstrations across 9 tasks, and validates the approach on a real-robot can-sorting task using a real-to-sim-to-real pipeline.

ExampleStarting from 60 human teleoperated demonstrations, the system automatically generates about 21,000 bimanual dexterous-hand demonstrations in simulators such as robosuite, which are then used to train a policy with behavioral cloning.

Also called

DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning

Related

MimicGen · DemoGen · Bimanual Manipulation · Real-to-Sim-to-Real · Synthetic Data · Behavior Cloning

Sources

DexMimicGen 项目主页 (Chinese)
DexMimicGen (arXiv)

As of

2024-10