HumanPlus
Stanford University · 2024 · datasets.bot · datasets.bot page
One-liner. Stanford CoRL 2024 system + HDF5 imitation-learning demos on a 33-DoF Unitree H1 humanoid for whole-body skills via human shadowing.
Setup
- Datasets / benchmarks: HumanPlus is a Stanford full-stack system that lets a customized 33-DoF Unitree H1 humanoid shadow human body and hand motion in real time from RGB cameras, and then learn autonomous whole-body skills via behavior cloning on teleoperated demonstrations. The low-level shadowing policy is trained in simulation using the 40-hour AMASS human-motion dataset; the released task data consists of HDF5 imitation-learning episodes (ACT/Mobile-ALOHA style) recording two head-mounted egocentric RGB cameras plus 19-DoF body and dexterous-hand joint positions. Demonstrated skills include folding clothes, rearranging objects, warehouse unloading, two-robot greeting, wearing a shoe, and typing. License: unknown. Download: https://drive.google.com/drive/folders/1i3eGTd9Nl_tSieoE0grxuKqUAumBr2EV.
- Hardware / simulator: Embodiment: human, unitree_humanoid. Environment: lab, warehouse, tabletop. Realness: physical.
Schema
episodes (HDF5) -> timesteps -> {observation: 2x egocentric RGB images + proprioception (joint positions), action: target body+hand joint positions}
Links