MimicGen
NVIDIA (with The University of Texas at Austin) · 2023 · datasets.bot · datasets.bot page
One-liner. Sim manipulation datasets auto-generated by NVIDIA's MimicGen from ~200 human demos: 48K+ demos, 12 tasks, 4 arms, HDF5.
Setup
- Datasets / benchmarks: MimicGen is a data-generation system (CoRL 2023, NVIDIA + UT Austin) that synthesizes large-scale robot manipulation datasets by adapting a small set of human teleoperation demonstrations to new object poses, objects, and robot embodiments. The public release contains over 48,000 simulated task demonstrations across 12 tasks (categories: source/core/object/robot/large_interpolation) in robosuite/MuJoCo, stored as robomimic-compatible HDF5 with low-dim and image observations. The broader system generated 50K+ demos across 18 tasks over two simulators (robosuite, Isaac Gym) and a physical robot arm. License: CC-BY-4.0. Download: https://huggingface.co/datasets/amandlek/mimicgen_datasets.
- Hardware / simulator: Embodiment: franka_panda, kuka_iiwa, sawyer, ur5. Environment: simulation, tabletop. Realness: simulated.
Schema
HDF5 (robomimic format): per-task file -> demos -> per-step {obs (low-dim states + RGB images), actions, rewards/dones}
Links