Human Video Scaling in Robotics: What GR00T 2, DYNA-2, EgoScale and GEN-0 Mean for Data Buyers
Between late 2025 and November 2026, robot foundation models crossed a threshold: the largest training corpora stopped being robot data and became human data. Companies now report hundreds of thousands to over a million hours of egocentric human video, plus tens of thousands of hours of handheld UMI demonstrations, with robot teleoperation as a smaller, specialized layer. Here is what was announced, what the evidence does and does not show, and what it means if you buy or collect data.
TL;DR: Robotics is finding its scaling axis in human data. GR00T 2 pretrains on 402,730 hours of egocentric human video plus UMI and teleoperation data; DYNA-2 reports over 1 million hours of human video; EgoScale and DreamDojo build on tens of thousands of hours; GEN-0 reports power-law scaling on 270,000+ hours of real-world manipulation data. For data buyers, the bottleneck is shifting from hours to diversity, capture standards and action-relevant signal.
Key takeaways
- Human video is now the largest pretraining layer in several 2026 robot models, according to their developers.
- Several companies report predictable improvement as human or real-world data grows, though results come from their own evaluations.
- Robot data has not disappeared; it is used for embodiment grounding, post-training and evaluation.
- Handheld UMI and wearable-hand data form a growing middle layer between human video and teleoperation.
- Data value now depends on diversity, capture consistency, rights and the signals each hour carries, not raw hour count.
Timeline: the human-data turn in robotics
| Date | Organization / model | Reported data | Why it matters |
|---|---|---|---|
| Mar 2025 | NVIDIA GR00T N1 | "Data pyramid" of web and human video, synthetic data and real robot data | Formalized a layered data strategy with latent actions for human video |
| Mid-2025 | Tesla Optimus | Shift toward worker-worn camera rigs, per Business Insider | A major humanoid program moving collection to human video |
| Sep 18, 2025 | Figure Project Go-Big | Egocentric human video in real homes | Reported humanoid navigation learned from human video only |
| Nov 4, 2025 | Generalist GEN-0 | 270,000+ hours of real-world manipulation data, growing 10,000+ hours per week | Reported power-law scaling with pretraining data |
| Nov 19, 2025 | Sunday ACT-1 | Skill Capture Glove data, no teleoperation | Wearable data matched to the robot hand |
| Dec 2025 | Physical Intelligence | Egocentric human video co-trained with π0.5 | Human-to-robot transfer emerges with diverse robot pretraining |
| Feb 2026 | NVIDIA EgoScale | 20,854 hours of action-labeled egocentric human video | Large-scale recipe for dexterous manipulation from human data |
| Feb 2026 | NVIDIA DreamDojo | 44,711 hours of human egocentric video | Robot world model pretrained on human video |
| Aug 10, 2026 | Dyna Robotics DYNA-2 | Over 1 million hours of egocentric human video | Claims a human-to-robot scaling law with no robot data in pretraining |
| Nov 2026 | NVIDIA GR00T 2 | 402,730 h human video; 10,184 h UMI across 2,897 tasks; 23,500 h teleoperation | Transfer from humans, UMI and DexUMI to robots that never practiced the task |
What the evidence shows
More human data improves robot performance, in these setups
NVIDIA's GR00T 2 provides one of the clearest controlled measurements: pretraining four models on none, 1%, 10% and 100% of its human video, real-robot task progress across 72 tasks rose from 50% to 70%. Generalist's GEN-0 reports a power-law relationship between pretraining data scale and downstream post-training performance, and a model-size threshold around 7B parameters below which models struggle to absorb the data. Dyna's DYNA-2 announcement claims performance improves smoothly as human video hours grow.
Transfer without task-specific robot demonstrations
GR00T 2 reports tasks shown only by humans, only with a handheld UMI gripper, or only with the DexUMI wearable hand being performed by robots that never practiced them. Physical Intelligence found human-to-robot transfer emerging once its VLA had diverse robot pretraining. Figure reported navigation learned from egocentric human video alone.
World-action models change what video is for
Both GR00T 2 and DYNA-2 are described as world-action models that predict future video together with actions. That design lets a model learn physics from how scenes evolve in human video, even when the video has no robot action labels. It is a key reason plain egocentric video has become so valuable.
What the evidence does not show
- Independent verification. Most figures come from the developers' own evaluations and announcements.
- That robot data is obsolete. GR00T 2 still uses 23,500 hours of teleoperation and reports large gains from a handful of robot demonstrations per task.
- That any video works. Recipes specify capture, labels and alignment. EgoScale's video is action-labeled; PI uses 3D hand positions as actions.
- Comparable hour counts. An hour of daily-life video, an hour of manipulation video with 3D hand pose, and an hour of UMI data carry very different signals.
What this means for data buyers and collectors
| Shift | Implication for collection |
|---|---|
| Human video is the largest layer | Plan for scale: contributor networks, light rigs, automated QA. |
| Diversity drives generalization | Budget for people, regions, homes and workplaces, not repeated scenes. |
| 3D hand signals become action labels | Add stereo, device tracking or wrist views to a subset of collection. |
| Handheld data is a growing middle layer | Collect UMI-style or wearable-hand data matched to target end effectors. |
| Multi-view robot inputs | Match head and wrist camera layouts to the robots you train. |
| Commercial models need rights | Consent, permitted-use terms and provenance metadata on every episode. |
The new bottleneck: not hours of video, but hours of the right video: diverse, consistently captured, legally usable and carrying the signals your model learns from.
Practical recommendations for 2027 planning
- Separate layers in your data plan: broad egocentric video, action-rich human data (stereo, wrist cams, gloves), handheld UMI, and robot data, each with its own target and budget.
- Standardize capture specs early. Consistent devices, frame rates, mounts and calibration make hundreds of thousands of hours usable together.
- Measure diversity explicitly. Track unique contributors, environments, objects and tasks, not just hours.
- Keep provenance. Device, calibration, consent and annotation versions per episode.
- Pilot against your model. Train on a small batch and check whether it moves your metrics before scaling.
Field Motion's role
Field Motion collects egocentric human data for robotics and physical AI (mono, stereo, multi-camera and UMI-style) through a contributor network in Latin America, Southeast Asia and the United States. That is the layer of the 2026 data stack that scales with people rather than robots. Explore the comparisons in this series: stereo vs. mono, UMI vs. teleoperation vs. human video, multicam vs. head-mounted, and public egocentric datasets. Earlier context is in Physical AI data shifts in 2026. Or visit the Field Motion homepage.
Frequently asked questions
Why are robot foundation models training on human video in 2026?
Human video scales far beyond robot data because it needs only people and cameras, not robots. Developers such as NVIDIA, Dyna Robotics and Physical Intelligence report that adding human video improves robot performance and enables transfer to tasks the robot never practiced, especially with world-action model architectures.
How much human video does GR00T 2 use?
NVIDIA reports that GR00T 2 is pretrained on 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation.
Is there a scaling law for robotics?
Several companies report scaling behavior. Generalist reports a power law between pretraining data and downstream performance for GEN-0, and Dyna Robotics claims a human-to-robot scaling law for DYNA-2. These results come from the companies' own evaluations.
Does human video replace robot teleoperation data?
Not entirely. Even human-video-heavy models such as GR00T 2 use robot teleoperation for embodiment grounding and adapt quickly with a few robot demonstrations per task. Human data reduces how much robot data is needed.
How does Field Motion fit into the human-data trend?
Field Motion collects egocentric human data (mono, stereo, multi-camera and UMI-style) through contributors in Latin America, Southeast Asia and the United States, supplying the diverse human layer that 2026 robot models scale on.
Building your 2027 data plan?
We'll help you size the human-video, action-rich and handheld layers and start with a pilot that proves value on your model.
Talk to Field MotionRelated reading
Sources
- NVIDIA — GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (March 2025)
- Business Insider — Inside the Strategy Shift at Optimus, Tesla's Humanoid Robot Program (August 2025)
- Figure — Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer (September 18, 2025)
- Generalist AI — GEN-0: Embodied Foundation Models That Scale with Physical Interaction (November 4, 2025)
- Sunday Robotics — ACT-1: A Robot Foundation Model Trained on Zero Robot Data (November 19, 2025)
- Physical Intelligence — Emergence of Human to Robot Transfer in Vision-Language-Action Models (December 2025)
- NVIDIA — EgoScale (February 19, 2026)
- NVIDIA — DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos (February 2026)
- Dyna Robotics — DYNA-2 World-Action Model announcement (August 10, 2026)
- NVIDIA GEAR — GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation (November 2026)
Published by Field Motion on November 9, 2026. Figures are as reported by the cited organizations, many from their own evaluations. Field Motion is not affiliated with the cited projects.