Glossary
What Is a Humanoid Robot?
Last reviewed: September 2026
Short answer
A humanoid robot is a robot built with a human-like body plan — a head, torso, two arms with hands, and usually two legs — so it can operate in spaces and use tools built for people. That body plan is why egocentric human video, captured from a head-worn viewpoint with both hands in frame, closely matches its training data.
01
Why it matters
A humanoid’s body plan is what makes egocentric human demonstration data transferable at all: a head-mounted camera occupies roughly the viewpoint a humanoid’s own head camera sees from, and two human hands map onto two robot end-effectors in a way that a single-arm manipulator or a wheeled mobile robot does not. That is also why humanoid programs lean on human video for scale. NVIDIA’s GR00T N1 describes training a humanoid foundation model on a data pyramid that widens from a small layer of real robot trajectories up through synthetic data to a broad base of web and human video, and Figure has publicly described Project Go-Big as a large-scale data collection effort aimed at its humanoid platform. But the body-plan match is not a perfect one: a human hand is not a robot end-effector, and human proportions, reach, and joint limits differ from a given humanoid’s hardware — the gap this creates between the data a model trains on and the robot it deploys to is called the embodiment gap, and it is the main reason humanoid data programs plan for a mid-training or alignment step rather than treating human video as a drop-in substitute for robot data.
02
How Firsthand handles it
Firsthand captures custom training data for humanoid robots as synchronized egocentric video recorded from a head-worn viewpoint, with two wrist cameras, depth, an IMU, and 3D hand and body pose for both hands — ten streams aligned under 2 ms per episode. Capture is scoped to a written spec covering tasks, environments, and accept criteria, billed only for validated hours, and delivered in RLDS, LeRobot format, or WebDataset under a buyer-owned commercial license. Where coverage is already deep, a first validated batch lands in a median of 48 hours from signed spec; thinner environments typically take three to four weeks because vetting operators there is the bottleneck.
03
Humanoid robots vs. other robot form factors, at a glance
| Form factor | Typical action space | Data implication |
|---|---|---|
| Humanoid | Two arms and hands, head-mounted viewpoint, whole-body balance and locomotion | Egocentric human video from a head-worn rig is a close viewpoint and action match |
| Single-arm manipulator | One arm, fixed base, third-person or wrist-mounted view | Teleoperation or third-person demonstration on the specific arm typically fits better than egocentric human video |
| Mobile manipulator | One or two arms on a wheeled base, no whole-body balance requirement | Navigation and manipulation data are often collected and specified separately |
| Quadruped | Four-legged locomotion, no hands or arms in most configurations | Human egocentric data has little action-space overlap; locomotion data comes from other sources |
05
Humanoid data by training stage
| Stage | Goal | Typical data |
|---|---|---|
| Pretraining | Broad bimanual and whole-body coverage across tasks, scenes, and people | Large-scale egocentric human video, e.g. the base layer of NVIDIA’s GR00T N1 data pyramid |
| Mid-training | Align human and robot embodiments on matched tasks and viewpoints | A smaller, tightly aligned human-plus-robot slice, e.g. EgoScale’s ≈50h human + ≈4h robot across 344 tabletop tasks |
| Post-training (fine-tuning) | Specialize the model to a specific deployment task on the target humanoid hardware | A written skill spec collected against the target task and edge cases |
06
Frequently asked questions
- What is a humanoid robot?
- A humanoid robot is a robot built with a human-like body plan — a head, torso, two arms with hands, and usually two legs — so it can operate in spaces and use tools designed for people, rather than requiring a purpose-built environment.
- How is a humanoid robot different from other robots?
- Most other robot form factors depart from the human body plan: a single-arm manipulator has one fixed arm and no legs, a mobile manipulator pairs one or two arms with a wheeled base, and a quadruped has four legs and typically no arms or hands. A humanoid keeps the two-arm, head-mounted, whole-body layout, which is what lets egocentric human data transfer to it more directly than to the other form factors.
- What training data do humanoid robots need?
- Humanoid foundation models need data that captures a head-mounted viewpoint, both hands in frame, and whole-body pose together — not just an isolated hand or arm. NVIDIA’s GR00T N1 describes combining broad human-video pretraining with a smaller layer of real robot trajectories to cover this.
- Why is egocentric human video useful for humanoid robots?
- Because a humanoid perceives from a head-mounted camera and acts with two hands, egocentric human video recorded from the same head-worn viewpoint is a closer match to its observation and action space than third-person video or single-arm teleoperation, which is why it scales well for pretraining-stage coverage.
- How do pretraining, mid-training, and post-training apply to humanoids?
- Pretraining uses broad, large-scale egocentric human video for general bimanual and whole-body coverage. Mid-training uses a smaller, tightly aligned human-plus-robot slice to bridge the embodiment gap — EgoScale’s published recipe pairs roughly 50 hours of human video with about 4 hours of robot data across 344 tabletop tasks. Post-training (fine-tuning) specializes the model to one task on the target hardware using a written skill spec.
- What is the embodiment gap for humanoid robots?
- The embodiment gap is the difference between a human demonstrator’s body and a humanoid’s hardware: proportions, reach, joint limits, and hand geometry differ, so human action data does not transfer to a robot policy without an alignment or retargeting step.
- How do you collect custom humanoid robot training data?
- By writing a spec that covers tasks, environments, and accept criteria for the target humanoid embodiment, then capturing synchronized egocentric video, depth, IMU, and 3D hand and body pose from a head-worn rig against that spec, billed per validated hour and delivered under a buyer-owned license.
- Do humanoid robots always have legs?
- Most humanoid platforms include two legs for locomotion and whole-body balance, which is part of what distinguishes the form factor from a stationary or wheeled manipulator, though the defining feature is the human-like head, torso, and two-arm layout rather than legs specifically.
Related terms