Humanoid Robot Datasets
Curated datasets for humanoid robot locomotion, whole-body manipulation, and loco-manipulation. From simulated MuJoCo walking to real-world Unitree G1 teleoperation — the data you need for humanoid policy training.
Key Humanoid Datasets
| Dataset | Focus | Robot / Sim | Scale |
|---|---|---|---|
| HumanoidBench | Locomotion + manipulation | MuJoCo humanoid | 15 tasks, RL baselines |
| Unitree G1 Demos | Whole-body teleop | Unitree G1 | Growing collection |
| H1 Walking | Bipedal locomotion | Unitree H1 | Sim + real transfer |
| NVIDIA Isaac Humanoid | Sim locomotion | Isaac Sim humanoids | Millions of steps |
| CMU MoCap (retargeted) | Motion retargeting | Various humanoids | 2,500+ motion clips |
Challenges in Humanoid Data
Humanoid robots present unique data challenges compared to tabletop arms. The action space is high-dimensional (20–40+ joints), balance constraints are critical, and failures (falls) can damage hardware. Most humanoid training data comes from simulation with sim-to-real transfer, but the gap between simulated and real-world dynamics remains significant for contact-rich tasks.
Emerging approaches use VR teleoperation with motion retargeting to collect real-world humanoid demonstrations. SVRC supports this workflow with Unitree G1 and H1 platforms and full-body motion capture retargeting pipelines.
Humanoid Hardware at SVRC
We operate Unitree G1 humanoid robots for data collection and policy deployment. G1 features 23 DoF, dexterous hands, and ROS 2 integration. Available for purchase, lease, or managed data collection campaigns.
For locomotion-focused research, we also work with the Unitree H1 and support sim-to-real pipelines using Isaac Sim and MuJoCo.