DexMimicGen

NVIDIA Research (with UT Austin and UC San Diego) · 2024 · datasets.bot · datasets.bot page

One-liner. NVIDIA system that synthesizes ~21K bimanual dexterous sim demos from 60 human demos across 9 tasks; real GR-1 transfer.

Setup

Datasets / benchmarks: DexMimicGen is an automated data-generation system that synthesizes large-scale trajectory datasets for bimanual dexterous manipulation from a handful of human teleoperation demonstrations. It generated over 21,000 demonstrations from 60 source human demos across 9 simulation tasks in robosuite/MuJoCo, spanning three embodiments (bimanual Panda arms with parallel-jaw grippers, bimanual Panda arms with dexterous hands, and a Fourier GR-1 humanoid with Inspire dexterous hands), and includes a Real2Sim2Real pipeline validated on a real GR-1 robot. License: CC-BY-NC-4.0. Download: https://huggingface.co/datasets/MimicGen/dexmimicgen_datasets/tree/main.
Hardware / simulator: Embodiment: franka_panda, humanoid_other, parallel_jaw. Environment: lab, simulation, tabletop. Realness: both.

HDF5 demo files per task -> episodes -> timesteps -> {observation (RGB-D, proprioception), action}