PHUMA — Physically-Grounded Humanoid Locomotion Dataset
PHUMA is a physically-grounded humanoid locomotion dataset developed by DAVIAN Robotics for training bipedal locomotion policies on Unitree G1 and H1_2 humanoid platforms. Released in 2025 under the Apache 2.0 license, the dataset contains retargeted human motion data with physics constraints applied to ensure physical feasibility on real robot hardware. It covers 100,000 to one million motion samples stored in NumPy format on Hugging Face. Unlike direct teleoperation datasets, PHUMA uses human motion capture as a source and retargets trajectories to the kinematic structure of Unitree humanoids, enabling diverse and natural locomotion policies that generalise across varied terrain. The dataset addresses a significant gap in publicly available locomotion training data for production humanoid platforms. Both Unitree G1 and H1_2 are priority embodiments in the Geppetto robot directory, and their supply chains are documented across reducers, motors, sensors, and screw components through the Geppetto Hardware Supply Chain database.
| Year | 2025 |
|---|---|
| Embodiments | Unitree G1, Unitree H1_2 |
| Modalities | proprioception, imu |
| Task categories | locomotion |
| Data format | npy |
| License | Apache 2.0 |
| Access | open — commercial use permitted |
| Maintainer | DAVIAN Robotics |
| Origin country | KR |
What is it?
PHUMA (Physically-Grounded Humanoid Locomotion Dataset) is a human motion capture dataset retargeted to Unitree G1 and H1_2 humanoid robot kinematics, developed by DAVIAN Robotics and released in 2025 under the Apache 2.0 license. It provides physically-constrained locomotion trajectories derived from human demonstrations — raw human motion is filtered through a physics simulator to ensure each trajectory is executable on real robot hardware without violating joint limits or dynamic feasibility constraints. The result is a large-scale source of diverse, natural locomotion policies that generalise across varied terrain types.
Who is it for?
PHUMA is designed for robotics researchers working on bipedal locomotion policy training for Unitree G1 and H1_2 humanoid platforms. It is particularly valuable for teams that want diverse, natural locomotion data — the human motion capture source provides far more variety than data collected directly from robots — but need physically feasible trajectories rather than raw kinematic retargeting. The Apache 2.0 license makes it suitable for commercial robotics development.
Key specifications
- Embodiments: Unitree G1, Unitree H1_2
- Task category: Locomotion (bipedal walking, varied terrain and gait patterns)
- Format: NumPy (.npy) — available on Hugging Face
- License: Apache 2.0 — commercial use permitted
- Access: Open — freely downloadable via Hugging Face
- Origin: South Korea (DAVIAN Robotics)
How it compares
PHUMA is distinctive in its methodology: it starts from human motion capture (diverse and natural) and applies physics constraints (ensuring robot feasibility), rather than collecting data directly from robots (limited variety) or using pure simulation (potential sim-to-real gap). This positions it between pure teleoperation datasets like GR00T Teleop G1 and simulation datasets. For locomotion specifically on Unitree G1 and H1_2, PHUMA is the primary open dataset available as of 2026.
Limitations and access notes
PHUMA covers locomotion only — it does not include arm manipulation, object interaction, or whole-body coordination tasks. The dataset targets Unitree G1 and H1_2 specifically; retargeting to other humanoid platforms requires additional work. The physics-grounding process may filter out some expressive human motions that are not feasible for the target robot's joint limits. Apache 2.0 permits commercial use and modification with no share-alike requirement.
Linked professions
Frequently asked questions
What makes PHUMA different from other locomotion datasets?
PHUMA combines human motion capture (which provides diverse, natural movement variety) with physics-based feasibility filtering (which ensures each trajectory can be executed by the real robot without violating joint limits or dynamic constraints). Most locomotion datasets are either collected directly from robots (limited variety) or from simulation (potential sim-to-real gap). PHUMA bridges both approaches.
Which humanoid robots does PHUMA support?
PHUMA is designed for the Unitree G1 and Unitree H1_2 humanoid robots. Both are commercially available bipedal humanoid platforms from Unitree Robotics. The dataset includes kinematic retargeting to both platforms' joint structures.
Can PHUMA be used commercially?
Yes. PHUMA is licensed under Apache 2.0, which permits commercial use, modification, and distribution. There is no share-alike requirement.
How do I access PHUMA?
PHUMA is freely available on Hugging Face at huggingface.co/datasets/DAVIAN-Robotics/PHUMA. The dataset is distributed in NumPy (.npy) format. No registration is required.
Does PHUMA include manipulation tasks?
No. PHUMA covers locomotion only — bipedal walking, gait variation, and terrain traversal. It does not include arm manipulation, object grasping, or whole-body coordination. For manipulation data on the Unitree G1, see the PhysicalAI GR00T Teleop G1 dataset.