How Much Training Data Does a Humanoid Robot Need?

No. Every publicly documented humanoid or robot foundation model effort states a different scale, because the teams behind them weighted hours, diversity, and embodiment alignment differently, and because none of them is answering the same question. EgoVerse reports roughly 1,362 hours across about 80,000 episodes. Egocentric-10K totals around 10,000 hours. EgoScale pretrains on 20,854 hours. GEN-0 draws on more than 270,000 hours. Figure’s Project Go-Big, a humanoid-specific initiative, publishes no hour count at all — only a property-access scale through its Brookfield partnership.

That three-order-of-magnitude spread, from one humanoid-relevant program to the next, is itself the answer to "how much data does a humanoid need": enough depends entirely on what question the data is meant to answer — a general pretraining prior, a narrower mid-training alignment step, or a specific fine-tuning task — and how diverse and well-labeled the hours that were captured actually are.

The clearest way to see the range is side by side, with each figure taken exactly as published by its own official source:

These are the only hour counts and scale statements used on this page, each taken exactly as published by its official source and linked in Sources below. No hour count, throughput figure, or customer name for Firsthand-collected data is implied or stated by this table.

Two of these six are humanoid-specific in their framing: Figure’s Project Go-Big is explicitly a humanoid pretraining initiative for Figure’s Helix model, and GR00T N1 is NVIDIA’s humanoid foundation model paper. The other four — EgoVerse, Egocentric-10K, EgoScale, and GEN-0 — are general robot or egocentric-pretraining corpora that a humanoid program can draw on, but were not built exclusively for humanoid embodiments. Mixing the two categories without noting that distinction is a common source of confusion when comparing "how much data" across projects.

A humanoid foundation model is not trained in one pass. This site’s pretraining vs mid-training vs post-training breakdown describes three distinct stages, each with a different data goal, and "how much data" only means something once you know which stage you are asking about:

Pretraining data, by far the largest of the three in published corpora, is judged on breadth: how many homes, workplaces, task categories, objects, and people it spans, so the model does not overfit to one narrow setting. Mid-training data narrows that broad prior toward a specific humanoid’s embodiment and deployment domains, while remaining more diverse than a fine-tuning batch. Fine-tuning data is the smallest and most tightly scoped of the three, built to represent one exact task as precisely as possible rather than to cover breadth.

See /guides/pretraining-vs-post-training-robot-data for the full three-stage comparison, including how each stage’s data is sourced and validated.

A humanoid hand with many degrees of freedom has a far larger action space than a parallel-jaw gripper or a simple two-finger end effector. EgoScale’s published scaling-law result makes this concrete: the paper measured a 54% improvement in task-completion success on a 22-degree-of-freedom dexterous hand when scaling from a smaller to the full 20,854-hour egocentric pretraining corpus (arXiv 2602.16710, NVIDIA GEAR). That is a direct, published data point for how much a high-DoF hand benefits from additional pretraining scale compared to a lower-DoF end effector.

The practical implication is that a humanoid hand’s dexterity is itself a reason to weight data collection toward bimanual, hand-pose-rich capture rather than assuming the same volume used for a simpler gripper would transfer. A policy that must coordinate many finger joints across two hands has more to learn per task than one that opens and closes a single gripper, which is part of why egocentric human video — which naturally captures full hand articulation — is a better pretraining match for a humanoid than teleoperated data from a simpler robot arm.

The published evidence says diversity and validated, labeled hours matter at least as much as raw volume, not that volume is irrelevant. EgoScale’s scaling-law study reports a log-linear relationship between pretraining-data volume and downstream performance: task-completion success improved roughly log-linearly with data scale, which means early hours in an underrepresented condition move performance more than additional hours in an already well-covered one — a diversity argument, not a pure volume argument. GEN-0, at more than 270,000 hours, reports on its official blog that scaling was accompanied by an emphasis on data quality and diversity, not sheer hour count alone (Generalist AI).

• Diversity of scenes, tasks, objects and people — a corpus concentrated on a few environments underperforms a smaller, more varied one
• Validated hours, not raw recording — footage that passed QA against a written spec, not everything captured
• Action labels — hand pose, wrist trajectories and segment boundaries that make video usable for training, not just viewable
• Calibration and sync quality — matched intrinsics and time-aligned streams that make geometry and depth usable across a corpus

See /guides/robot-data-scaling-laws for the full EgoScale and GEN-0 scaling-law discussion.

NVIDIA’s GR00T N1 paper (arXiv 2503.14734) describes training its humanoid foundation model on a pyramid of data sources at different scales and levels of realism: broad web and human video data at the base, narrowing through latent-action and inverse-dynamics-model (IDM) pseudo-labels, up to real robot trajectory data at the top of the pyramid, which is the smallest and most expensive tier to collect. Egocentric human video sits in the middle of that pyramid — more physically grounded than general web video, far more scalable to collect than real robot teleoperation.

Figure’s Project Go-Big is a published data point for just how far the human-video side of that pyramid can carry a humanoid model: Figure states that training its Helix model on 100% egocentric human video, with no robot demonstration data at all, produced the company’s first zero-shot human-to-robot navigation transfer — Helix responding to commands like "walk to the kitchen table" by generating closed-loop navigation control through cluttered home environments, which Figure describes as the first time a humanoid has learned this end-to-end from human video with no robot-specific training (figure.ai/news/project-go-big). Figure’s own announcement is explicit that this result is scoped to navigation, not manipulation.

Read together, GR00T N1’s pyramid and Project Go-Big’s reported navigation result both point the same direction: egocentric human video is not a stand-in to be minimized until "real" robot data arrives, but a legitimate, scalable pretraining tier in its own right, with real robot trajectory data reserved for the narrower top of the pyramid where embodiment-specific precision matters most.

None of the public reference points above settle how much data any one specific humanoid program needs, because each was built for a different model, a different embodiment, and a different evaluation setup. A program scoping its own humanoid data collection should start from its own coverage gaps rather than matching any single announced project’s hour count.

A written skill spec sets targets for environments, task categories, object sets, and contributor demographics before capture begins, so the resulting corpus does not overfit to one room, one task, or one body type. That spec becomes the basis for a quote on reach, timeline, and price, and the same spec is what every batch is validated against before it ships.

• Define the training stage first — pretraining breadth, mid-training alignment, or fine-tuning precision — since each has a different volume and diversity target
• Write a spec that sets explicit environment, task, and contributor coverage targets, not just a target hour count
• Specify the embodiment match needed: head-worn viewpoint, bimanual hand pose, whole-body pose and IMU for locomotion-to-manipulation transfer
• Run a small pilot batch scored against the spec before committing to full volume

See /custom-collection/humanoid-pretraining-data for how this site describes bimanual, whole-body egocentric capture planned for diversity across homes, workplaces, tasks, and people, and /guides/robot-pretraining-data for how pretraining data differs from fine-tuning data generally.

Many humanoid programs start with a Spec Pilot — a small, fast first batch that proves coverage and label quality against the written spec before scaling to full volume. Firsthand runs custom humanoid collections through the same seven-stage pipeline used for every collection on this site, with a median 48-hour first validated batch where coverage is deep, and a longer timeline where contributors need to be recruited first because coverage is thin.

Every custom humanoid collection ships under a perpetual, buyer-owned commercial license, with faces, plates, and screens irreversibly blurred at ingest and a documented consent chain behind every episode — the same terms stated across every Firsthand custom collection, not a humanoid-specific exception.