What Is a Humanoid Foundation Model?

A humanoid foundation model is only as general as the data under it. The GR00T N1 paper describes a pyramid: abundant web and human video at the base, synthetic data in the middle, scarce real robot data at the top. That is why human video matters for humanoids: it is the cheapest way to add breadth, while robot data stays reserved for the embodiment-specific top.

Firsthand collects data for the human-video part of the pyramid; it does not build or train models. Capture is synchronized egocentric video 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. It is scoped to a written spec, billed only for validated hours, and delivered in RLDS, LeRobot format, or WebDataset under a buyer-owned commercial license.

It is a single large model meant to drive many tasks on a humanoid body, rather than one policy per task. “Generalist” describes that breadth. The model is pretrained once on broad data, then fine-tuned or aligned to the robot it will run on.

The term is used in the GR00T N1 paper, which describes a generalist humanoid robot foundation model. This page covers the data side of the idea, not any one model’s architecture, which the site does not state.

A robot foundation model is the wider category: a large policy pretrained on broad robot-learning data, sometimes across embodiments. A humanoid foundation model targets one body plan, with a head, two arms with hands, and usually two legs.

That body plan is why egocentric human video transfers at all: a head-mounted camera occupies roughly the viewpoint a humanoid head camera sees from, and two human hands map onto two end-effectors. The match is imperfect, and the resulting embodiment gap is why programs plan an alignment step.

The GR00T N1 paper (arXiv 2503.14734) describes a mixture arranged like a pyramid by abundance and specificity. The table below summarizes it as this site describes the paper.

This site treats egocentric human video as more physically grounded than general web video and far more scalable to collect than real robot teleoperation. The paper’s hour counts are not stated here.

NVIDIA’s EgoScale pretrained a robot foundation model on 20,854 hours of action-labeled egocentric human video. Its labels came from off-the-shelf SLAM and hand-pose estimation after capture, and it reports a log-linear relationship between hours and downstream success with no observed saturation.

Its mid-training stage used about 50 hours of human video and about 4 hours of robot data across 344 tabletop tasks, evaluated on a 22-degree-of-freedom dexterous hand, with a reported 54% average success-rate improvement versus no pretraining. A humanoid evaluation is not stated.

HumanScale is published research on the scaling behavior of human-video pretraining for robot policies. Beyond that description, its figures are not stated here.

Firsthand delivers custom egocentric data, collected fast and on demand to a written spec. The second table separates what the site states from what it does not. Firsthand has no stated relationship with NVIDIA or any model builder, and this page describes public research only.

Start from your own coverage gaps rather than matching any announced project’s hour count, because each public result was built for a different model, embodiment and evaluation.

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.

What is a humanoid foundation model?

A large, generalist robot policy for a human-shaped body, pretrained on broad data and then adapted to a specific humanoid. The GR00T N1 paper describes one trained on a pyramid of web, human, synthetic and robot data.

What is a generalist humanoid model?

One model trained to handle many tasks on a humanoid rather than a separate policy for each task. The GR00T N1 paper uses the phrase generalist humanoid robot foundation model.

What training data does a humanoid foundation model need?

Per the GR00T N1 paper: web and human video at the broad base, synthetic data in the middle, and a small top layer of real robot trajectories matched to the deployed embodiment.

Does egocentric human video work for humanoid models?

This site treats it as a scalable, physically grounded layer, since a head-worn camera and two hands resemble a humanoid’s viewpoint and action space. The embodiment gap means programs still plan a mid-training alignment step.

How fast can custom humanoid data be collected?

Where coverage is deep, a first validated batch lands in a median of 48 hours from signed spec. Thinner environments typically take three to four weeks.

Does Firsthand build or train humanoid foundation models?

No. Firsthand collects and delivers custom egocentric data to a written spec. No relationship with NVIDIA or any model builder is stated, and no model is stated to be trained on Firsthand data.