Human2LocoMan — Loco-Manipulation Dataset
Human2LocoMan is a real-robot loco-manipulation dataset developed at Rniu Lab for training versatile quadrupedal robots capable of simultaneous locomotion and manipulation. Released in 2025 under the MIT license, the dataset contains demonstrations collected on the LocoMan platform and includes both human hand demonstration video and real robot trajectory data in a unified format designed for pretraining manipulation priors from human demonstrations. The dataset fills a critical gap in publicly available loco-manipulation training data — a task category requiring robots to walk and use arms simultaneously, which is foundational for real-world deployment in unstructured environments. Stored on Hugging Face with video modality in HDF5 and MP4 formats, the unified human-robot format enables efficient pretraining of imitation learning policies that transfer from human demonstrations to robot execution. Loco-manipulation is identified as a priority task gap in the Geppetto Datasets catalogue and this dataset is one of the first open contributions to this underserved category.
| Year | 2025 |
|---|---|
| Embodiments | LocoMan |
| Modalities | rgb, proprioception |
| View type | egocentric, wrist-cam |
| Task categories | loco-manipulation, manipulation |
| Data format | hdf5, mp4 |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | Rniu Lab (chrisyrniu) |
| Origin country | US |
What is it?
Human2LocoMan is a real-robot loco-manipulation dataset developed at Rniu Lab for training versatile quadrupedal robots capable of simultaneous locomotion and manipulation. Released in 2025 under the MIT license, the dataset contains demonstrations collected on the LocoMan platform and includes both human hand demonstration video and real robot trajectory data in a unified format. The methodology is designed for pretraining manipulation priors from human demonstrations that can then transfer to robot execution, targeting the underserved loco-manipulation task category — where robots must walk and use arms simultaneously.
Who is it for?
Human2LocoMan targets researchers working on loco-manipulation — one of the most challenging and least-resourced areas in robot learning, where the robot must coordinate locomotion and manipulation simultaneously. It is particularly valuable for researchers developing whole-body control policies and those studying how to leverage human demonstration data for robot learning without requiring direct robot teleoperation for all data collection.
Key specifications
- Robot platform: LocoMan (quadrupedal loco-manipulation platform)
- Data collection: Human hand demonstrations + real robot trajectories
- Task categories: Loco-manipulation, manipulation
- View type: Egocentric, wrist camera
- Format: HDF5, MP4
- License: MIT — commercial use permitted
- Access: Open — via Hugging Face
How it compares
Human2LocoMan fills a unique gap — most manipulation datasets use stationary arm platforms (DROID, EgoVerse, ALOHA) while most locomotion datasets have no manipulation component (PHUMA, GrandTour). Human2LocoMan is one of the only open datasets covering simultaneous locomotion and manipulation. The LocoMan platform is a quadruped with manipulation capability rather than a bipedal humanoid, which limits direct transfer to humanoid robots but makes the dataset valuable for whole-body coordination research.
Limitations and access notes
The LocoMan platform is a quadruped, not a bipedal humanoid — policies trained on this data may not transfer directly to humanoid platforms without adaptation. The dataset is small by modern standards; it is best used for fine-tuning or as a complement to larger datasets. MIT license permits unrestricted commercial use.
Linked professions
Frequently asked questions
What is loco-manipulation?
Loco-manipulation refers to tasks that require a robot to perform locomotion (moving through space) and manipulation (interacting with objects) simultaneously. For example, walking to an object and picking it up while continuing to move, or maintaining balance while using an arm to push a door open.
What robot does Human2LocoMan use?
Human2LocoMan uses the LocoMan platform, a quadrupedal robot with manipulation capability. Unlike bipedal humanoids, LocoMan uses four legs for locomotion and a separate manipulator for object interaction.
Can Human2LocoMan be used commercially?
Yes. Human2LocoMan is licensed under MIT, which permits commercial use, modification, and distribution with no restrictions.
How do I access Human2LocoMan?
Human2LocoMan is available on Hugging Face at huggingface.co/datasets/chrisyrniu/human2locoman. It is freely downloadable in HDF5 and MP4 formats. No registration is required.
Is Human2LocoMan useful for humanoid robot research?
Human2LocoMan is collected on a quadruped platform rather than a bipedal humanoid. The loco-manipulation coordination insights and human-to-robot transfer methodology are relevant to humanoid research, but direct policy transfer requires adaptation to bipedal kinematics.