The Dataset Viewer has been disabled on this dataset.

FreeAct · Data

Demonstration-Free Robot Adaptation
via Action-Grounded Generated Videos

NeurIPS 2026

Paper PDF  ·  Project Page  ·  Code  ·  Models

This repository contains the LIBERO demonstrations, generated videos and selected policy rollouts used to train FreeAct, in the LeRobot v3 format. DROID is hosted separately at Mo-ZheHan/droid.

Contents

Path Contents Episodes
libero/libero_spatial Source-domain demonstrations 432
libero/libero_{object,goal,10,90} Other LIBERO suites 5,182
generated/libero_{target,spatial}_wan2.6 Generated videos for stage-3 adaptation 100 each
generated/libero_{camera,bottle}_wan2.6 Generated videos of the two additional target environments (camera shift, ketchup bottle) 50 each
self_evolution/ Rollouts selected by the MLLM judge, one dataset per task 36

Format

  • Demonstrations and rollouts: 256 × 256 agent-view and wrist images at 20 fps; 8D state (end-effector position, axis-angle orientation, two gripper joint positions); 7D action (end-effector translation and rotation deltas, gripper command with 1 = open, 0 = closed).
  • Generated videos: 960 × 960 at 30 fps under observation.images.video, without action labels.

Sources

  • LIBERO demonstrations: LIBERO as regenerated by OpenVLA (no-op actions and failed replays removed): openvla/modified_libero_rlds (MIT) for libero_spatial, libero_object, libero_goal and libero_10, and IPEC-COMMUNITY/libero_90_no_noops_lerobot (Apache-2.0; 73 of the 90 tasks) for libero_90. We converted the IPEC-COMMUNITY LeRobot versions, whose gripper action is remapped from −1/+1 to 1/0, to LeRobot v3.
  • Generated videos: AI-generated by the authors with Wan 2.6 through the Alibaba Cloud Model Studio (Bailian) API, conditioned on a first frame rendered in the LIBERO simulator and a task instruction. They are synthetic and may be physically implausible; keep them labeled as AI-generated if you redistribute them.
  • Self-evolution rollouts: recorded in simulation by FreeAct policies and selected by an MLLM judge.

Download

hf download Mo-ZheHan/FreeAct-Data --repo-type dataset --local-dir data   # about 8 GB

Single datasets are downloaded with --include '<path>/*', e.g. --include 'libero/libero_spatial/*'. They can be loaded with LeRobotDataset(repo_id, root=..., video_backend="pyav").

Citation

@inproceedings{mo2026freeact,
  title     = {FreeAct: Demonstration-Free Robot Adaptation via Action-Grounded Generated Videos},
  author    = {Mo, Zhe-Han and Li, Jia-Ning and Song, Jiaxi and Guo, Meng-Hao and Hu, Shi-Min},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems},
  year      = {2026},
  url       = {https://openreview.net/forum?id=wNWxVVntG2}
}

Please also cite LIBERO and OpenVLA for the LIBERO demonstrations, and Wan for the generated videos:

@inproceedings{liu2023libero,
  title     = {{LIBERO}: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author    = {Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track},
  volume    = {36},
  pages     = {44776--44791},
  year      = {2023}
}

@inproceedings{kim2025openvla,
  title     = {{OpenVLA}: An Open-Source Vision-Language-Action Model},
  author    = {Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and Xiao, Ted and Balakrishna, Ashwin and Nair, Suraj and Rafailov, Rafael and Foster, Ethan P. and Sanketi, Pannag R. and Vuong, Quan and Kollar, Thomas and Burchfiel, Benjamin and Tedrake, Russ and Sadigh, Dorsa and Levine, Sergey and Liang, Percy and Finn, Chelsea},
  booktitle = {Proceedings of the 8th Conference on Robot Learning (CoRL)},
  series    = {Proceedings of Machine Learning Research},
  volume    = {270},
  pages     = {2679--2713},
  year      = {2025}
}

@article{wan2025wan,
  title   = {Wan: Open and Advanced Large-Scale Video Generative Models},
  author  = {{Team Wan} and Wang, Ang and Ai, Baole and Wen, Bin and others},
  journal = {arXiv preprint arXiv:2503.20314},
  year    = {2025}
}

License

MIT, except libero/libero_90, which is derived from IPEC-COMMUNITY/libero_90_no_noops_lerobot and released under the Apache-2.0 License. The LIBERO benchmark and OpenVLA's regenerated data are released under the MIT License.

Downloads last month
216

Models trained or fine-tuned on Mo-ZheHan/FreeAct-Data

Paper for Mo-ZheHan/FreeAct-Data