AutoData is a trajectory data-generation framework built on NVIDIA Isaac Lab and Isaac Lab-Arena. Given a small set of annotated human demonstrations, it uses parallel simulation environments to generate diverse datasets of successful robot demonstrations for robot learning.
AutoData splits demonstrations into object-relative skill segments. During generation, it transforms those segments to new scene configurations, connects them into complete trajectories, executes them in simulation, and records successful demonstrations in HDF5 datasets.
Robot-learning policies require large and diverse collections of successful demonstrations. Gathering all of that data through human teleoperation is slow and expensive, even though a small set of demonstrations often already contains the task's essential skills.
AutoData scales those demonstrations across randomized object placements and scene configurations, reducing the amount of manual collection needed to produce datasets for policy training.
- Demonstration amplification — Transform and recombine a small number of annotated demonstrations into larger, more diverse datasets.
- Three generation algorithms — Use MimicGen for single-arm tasks, DexMimicGen for coordinated multi-arm tasks, or SkillGen for collision-aware, motion-planned transitions.
- Parallel simulation — Generate demonstrations across multiple Isaac Lab environments at once.
- Robot and task configuration — Describe embodiments, tasks, subtasks, and generation policies in reusable YAML files.
- Adaptable environments — Apply environment profiles to create task variants without defining a new simulation environment.
- End-to-end dataset tools — Record, annotate, generate, validate, and replay HDF5 demonstrations.
- Linux with an NVIDIA GPU and driver that meet the Isaac Sim requirements
- Docker and the NVIDIA Container Toolkit
- Git and Git LFS
- An NGC account for the Isaac Sim container image
Clone the repository with its nested Isaac Lab-Arena and Isaac Lab submodules, then pull the example datasets:
git clone --recurse-submodules git@github.com:isaac-sim/AutoData.git
cd AutoData
git lfs install
git lfs pullLog in to NGC and launch the development container:
docker login nvcr.io
./docker/run_docker.shThe first launch builds the development image and opens a shell in the repository at
/workspaces/autodata. Subsequent launches reuse the image. Use ./docker/run_docker.sh -c to include cuRobo
for SkillGen workflows.
Docker is the recommended setup. An optional conda installation and additional container options are described in the installation guide.
Inside the container, generate ten Franka cube-stacking demonstrations from the pre-annotated dataset:
python scripts/generate_dataset.py \
--viz kit \
--env_name Isaac-Stack-Cube-Franka-IK-Rel-v0 \
--alg mimicgen \
--generation_num_trials 10 \
--num_envs 10 \
--task_descriptor autodata_examples/tasks/franka_cube_stack.yaml \
--embodiment autodata_examples/embodiments/franka_ik_rel.yaml \
--input_file ./datasets/annotated_datasets/dataset_franka_annotated.hdf5 \
--output_file ./datasets/generated_dataset_franka_quickstart.hdf5Validate the generated dataset:
python scripts/validate_dataset.py ./datasets/generated_dataset_franka_quickstart.hdf5See Your First Data Generation for an explanation of each step and instructions for replaying the result.
| Workflow | Description |
|---|---|
| Franka Cube Stacking | Record, annotate, and expand single-arm demonstrations with MimicGen. |
| Humanoid Pick-and-Place | Generate bimanual demonstrations for Fourier GR-1 and Unitree G1 with DexMimicGen. |
| Motion-Planned Generation | Use SkillGen and cuRobo to create collision-aware trajectories, including task variants built with environment profiles. |
AutoData/
├── autodata_core/ # Data generation algorithms and execution
├── autodata_interfaces/ # Task, embodiment, datastream, and planner interfaces
├── autodata_utils/ # Shared utilities
├── autodata_examples/ # Example task, embodiment, and environment-profile configs
├── autodata_tests/ # Unit, end-to-end, and performance tests
├── scripts/ # Dataset generation, annotation, and validation tools
├── datasets/ # Example and test datasets stored with Git LFS
├── docker/ # Reproducible development containers
├── docs/ # Sphinx documentation
└── submodules/ # Isaac Lab-Arena and Isaac Lab
Bug reports, feature suggestions, documentation improvements, and pull requests are welcome. Before opening a pull request, run the repository's pre-commit checks:
pre-commit run --all-filesFor test-suite details and common commands, see Testing and CI.
- Questions and ideas — GitHub Discussions
- Bug reports — GitHub Issues
- Isaac Sim questions — NVIDIA Developer Forums
AutoData is released under the Apache License 2.0.
Third-party open-source software notices are provided in THIRD_PARTY_NOTICES.md.
Additional component-specific licenses are stored in docs/licenses/. NVIDIA cuRobo is governed
by the NVIDIA Isaac Lab Additional Software and Materials License.
AutoData depends on Isaac Sim, which includes components distributed under proprietary licensing terms. See the Isaac Sim license for details.
If you use AutoData in your research, please cite:
@misc{autodata2026,
title = {AutoData: Scalable Robot Demonstration Generation for Robot Learning},
author = {{NVIDIA AutoData Contributors}},
year = {2026},
url = {https://github.com/isaac-sim/AutoData}
}Depending on the generation algorithm used, please also cite the original MimicGen, DexMimicGen, or SkillMimicGen work. Isaac Lab users should also cite the Isaac Lab paper.
AutoData builds on NVIDIA Isaac Sim, Isaac Lab, and Isaac Lab-Arena. Its data-generation workflows incorporate ideas from MimicGen, DexMimicGen, and SkillMimicGen, with cuRobo providing GPU-accelerated motion planning for SkillGen workflows.
We thank the authors and contributors of these projects, along with the broader robotics community, for their foundational work.
AutoData · Documentation · GitHub