Glossary
What Is HumanPlus?
Last reviewed: October 2026
Short answer
HumanPlus is a full-stack system from arXiv 2406.10454 for humanoids to learn motion and autonomous skills from human data. A simulation-trained policy lets a humanoid shadow a person’s body and hand motion using only an RGB camera, and teleoperated shadowing data trains egocentric-vision skill policies by behavior cloning. Firsthand is not part of it.
01
What is HumanPlus?
HumanPlus is introduced in arXiv 2406.10454, “HumanPlus: Humanoid Shadowing and Imitation from Humans”, with an official project page and GitHub repository. Read: the abstract, project page and README; not the full paper.
02
Where does Firsthand fit next to HumanPlus?
HumanPlus is third-party research. Firsthand is not affiliated with its authors and claims no use, training or compatibility with it or any named model. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec.
03
How does humanoid shadowing work?
Per the abstract, a low-level policy is trained in simulation by reinforcement learning on existing 40-hour human motion datasets, then transfers to the real world. It lets a humanoid follow human body and hand motion in real time using only an RGB camera. That is shadowing.
04
How does HumanPlus learn skills from human data?
Through shadowing, human operators teleoperate the humanoid to collect whole-body data in the real world. The abstract says supervised behavior cloning then trains skill policies using egocentric vision. The skill data is collected on the humanoid by teleoperation; training skills directly on human video is Not stated in the sources read.
05
What hardware and sensing does it use, as stated in the sources?
The abstract demonstrates the system on a customized 33-DoF, 180cm humanoid and says shadowing needs only an RGB camera. Camera models and resolution are Not stated in the sources read.
06
What did HumanPlus demonstrate?
The abstract lists six tasks, including folding a sweatshirt, typing and unloading objects from warehouse racks, with 60-100% success rates using up to 40 demonstrations. These are the authors’ own results; Firsthand draws no comparison.
07
How does HumanPlus relate to motion retargeting and loco-manipulation?
The sources read use neither term. The site defines motion retargeting as mapping human hand and wrist motion onto a robot with different kinematics; shadowing is a related idea, but that reading is Firsthand’s, not the authors’. Whether any listed task is loco-manipulation is Not stated in the sources read.
08
What does HumanPlus not describe?
Code and dataset licence terms, and the size of the dataset the README links, are Not stated in the sources read.
09
How do you get custom egocentric data for your own humanoid tasks?
In the sources, skill data is collected by teleoperating the humanoid. Firsthand captures human first-person video, not robot or teleoperation data. Write tasks, environments and signals into a spec and confirm them on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats: RLDS, LeRobot v2.1, WebDataset, HDF5; mapping to HumanPlus’ data layout is the buyer’s pipeline.
10
HumanPlus pipeline stages, as stated in the official sources
| Stage | What the official sources state |
|---|---|
| Low-level policy (abstract) | Trained in simulation by reinforcement learning on existing 40-hour human motion datasets; transfers to the real world |
| Shadowing (abstract) | Humanoid follows human body and hand motion in real time using only an RGB camera |
| Data collection (abstract) | Human operators teleoperate the humanoid through shadowing to collect whole-body data |
| Skill policies (abstract) | Supervised behavior cloning using egocentric vision |
| Named components (README) | HST for shadowing, HIT for imitation; WHAM (body) and HaMeR (hand) for pose estimation |
| Code and dataset licence | Not stated in the sources read |
12
HumanPlus beside related glossary terms
| Item | HumanPlus (sources read) | Motion retargeting (site entry) | Loco-manipulation (site entry) | Humanoid robot (site entry) |
|---|---|---|---|---|
| Kind | Full-stack shadowing and imitation system | Mapping human hand and wrist motion onto a robot | A robot moving its body while it manipulates | Robot with a human-like body plan |
| Term in the sources read | Not applicable | No | No | Yes: humanoids |
13
What HumanPlus consumes beside what the Firsthand site states
| Aspect | HumanPlus (sources read) | What the Firsthand site states |
|---|---|---|
| Viewpoint | Egocentric vision for skill policies | Head-mounted video plus two wrist cameras |
| Hands | Hand motion followed in real time | 21 keypoints per hand with a visibility flag, wrist trajectories |
| Body | Human body motion followed | 3D body pose; no head or camera 6-DoF pose, SLAM trajectory or whole-body motion capture stated |
| Segments and language | Not stated in the sources read | Action segments; one task instruction per episode |
| Robot actions | Collected by teleoperating the humanoid | Not captured: no robot, teleoperation or retargeted data |
| Compatibility | Not claimed | Not claimed |
14
Frequently asked questions
- What is HumanPlus?
- A full-stack system, introduced in arXiv 2406.10454, for humanoids to learn motion and autonomous skills from human data through shadowing and imitation.
- What is humanoid shadowing?
- A simulation-trained policy lets a humanoid follow human body and hand motion in real time using only an RGB camera.
- How does HumanPlus learn skills?
- Teleoperated shadowing collects whole-body data, then behavior cloning trains skill policies using egocentric vision.
- Is HumanPlus open source?
- The authors publish a GitHub repository. Licence terms are Not stated in the sources read.
- Does HumanPlus train skills on human egocentric video?
- Not stated in the sources read. Skill policies train on data collected by teleoperating the humanoid.
- Does Firsthand use HumanPlus, and can I request a custom dataset?
- Firsthand is not affiliated with the authors and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.
Related terms