One Million Datasets Later, Robotics Still Doesn't Have Enough Good Data

A dataset registry is a hosted repository where researchers and companies publish structured data — images, video, sensor logs, robot trajectories — for others to reuse in model training. Hugging Face, the platform most robotics and AI labs use to share that data, crossed a symbolic line this month: it now hosts more than one million public datasets, with NVIDIA counted among its largest single contributors of models and tooling built on top of them.

That number will get repeated in a hundred slide decks by year end. It shouldn't be treated as a proxy for progress. A registry with a million entries tells you almost nothing about how many of those datasets are usable for a specific robot, a specific camera mount, or a specific task — which is the only question that actually matters once you're trying to train a policy that works outside a demo reel.

This post walks through what the milestone actually signals, why egocentric and robot-training data has scaled faster than curation practices can keep up with in 2026, and what a field data collection team should check before treating any public dataset as a shortcut.

The Milestone, in Context

Hugging Face's dataset library is the default place robotics and AI teams look first when they need training data instead of collecting it themselves. This month it passed one million hosted datasets, and NVIDIA is already a major model contributor to Hugging Face, which recently surpassed 1 million datasets (The Robot Report, accessed September 7, 2026). That scale didn't happen in isolation. NVIDIA and Hugging Face have spent 2026 building tooling specifically to make dataset contribution frictionless: NVIDIA Isaac Teleop is an open-source framework for robot data collection that lets developers capture high-quality human demonstrations from external devices using standardized formats, then expand and share datasets with the community directly in LeRobot (The Robot Report, accessed September 7, 2026).

Lower friction to publish is good. It's also exactly why raw dataset counts stopped being a useful signal somewhere around dataset number 400,000. A registry doesn't check whether your kitchen footage matches a warehouse task, whether your camera height matches your robot's mount point, or whether your annotations were done by someone who understood the task at all.

Scale Has Never Been the Real Bottleneck

Egocentric video — first-person footage captured from a wearable camera or headset — has followed the same trajectory as the broader registry. Independent tracking of public releases shows a progression that is staggering: 10,000 hours in November 2025, 100,000 in December, 1 million in April 2026 (Labellerr, accessed September 7, 2026). That's not incremental growth. That's four orders of magnitude in five months, and it happened because collection got cheap, not because curation got easier.

The more useful research question isn't how much data exists — it's whether more data actually improves a robot policy. On that front, at least one 2026 study offers real evidence rather than a marketing claim: EgoScale trained a vision-language-action model on over 20,854 hours of action-labeled egocentric human video and demonstrated a log-linear scaling law between human data scale and validation loss, with that loss strongly correlating to downstream robot performance — the first published evidence connecting egocentric data volume directly to robot policy quality (Digital Divide Data, accessed September 7, 2026). That's a genuinely useful finding. It's also narrower than "more data is always better" — it says scale helps within a controlled, curated collection design, not that any million hours will do.

The same source is blunt about what separates a working dataset from a dumping ground: the best-performing egocentric programs sample across environments, objects, lighting conditions, and demonstrators systematically rather than collecting large volumes in a narrow setting . That's a collection design principle, not a scraping strategy. It's the difference between a program built around a research question and a program built around a hard drive quota — a distinction we've written about before in the context of how robot training data actually gets priced.

What "High Quality" Actually Requires on the Ground

Hardware vendors are already building around the assumption that generic footage isn't enough. At Automate 2026, Teradyne Robotics used Orbbec's Gemini 305g stereo 3D cameras within its UR AI Trainer, a platform for high-quality data collection required for sophisticated AI and VLA model training , and Orbbec reports it has served more than 1,600 robotics companies worldwide, achieving market shares of over 70% in both China's service robot vision market and South Korea's mobile robot vision market (The Robot Report, accessed September 7, 2026). That's a lot of specialized capital going into capturing data correctly the first time, which tells you something about how expensive it is to fix badly captured data after the fact.

Robotics data collection also can't be scraped the way text or image data can. As one recent infrastructure analysis put it plainly: Scale AI represents a foundational layer — the system required to turn raw physical interactions into usable training data — because robotics data cannot be collected from the internet the way text or images can (The Robot Report, accessed September 7, 2026). Every hour has to be produced, physically, by a person or a robot doing something in a real space.

Metric	Volume-first sourcing	Task-first field collection
Selection criterion	Total hours available in a public registry	Camera height, lighting, and task match to production environment
Usable rate	Often under 20% after filtering for mismatch	Designed to be usable at collection, not filtered after
Provenance	Frequently unclear who captured it or under what conditions	Documented at time of capture, tied to a specific research question
Cost driver	Storage and re-annotation of mismatched footage	Fieldwork design and demonstrator recruiting

The Market Research Parallel

Market research hit this exact wall years before robotics did, just with survey panels instead of egocentric video. Volume without governance produces expensive noise, not insight — which is why the industry's benchmarking work has shifted so hard toward fraud detection and provenance standards. Between 70% and 88% of professionals now use automated fraud detection, and firms with clear governance policies consistently outperform those without them (Veridata Insights, citing Greenbook's GRIT Insights Practice Report, accessed September 7, 2026). The 2026 edition of that same report found quality infrastructure still matters most, with trusted panels and fraud detection remaining essential in AI-driven research , and that mid-size research firms are now leading the industry in revenue growth, capability expansion, and AI governance maturity, while the largest firms are three years into a deliberate exit from fieldwork toward consulting and analytics (Greenbook, accessed September 7, 2026).

That last point matters more than it looks. The firms winning aren't the ones with the biggest panels — they're the ones with the tightest control over how the data got collected. Robotics is heading toward the same fork. Whoever controls collection methodology, not whoever hosts the most files, ends up owning the usable data. We've made a version of this argument before about verification gaps in research fieldwork in a piece on the distance between a desk and a verified field observation.

A Worked Example

Consider a warehouse robotics team evaluating whether to license a 40,000-hour public egocentric dataset instead of running its own capture program. The dataset was collected in kitchens and living rooms, at a headset height of roughly 1.6 meters, under indoor lighting that never drops below 300 lux. Their target task is a pick-and-place station with a wrist-mounted camera 40 centimeters off the ground, under warehouse lighting that swings from 150 to 800 lux depending on the loading dock doors. After filtering for camera angle, task category, and lighting range, the team typically keeps under 10% of the licensed hours — not because the dataset is fraudulent or poorly labeled, but because it was never built for their embodiment. The one million hours never mattered. The 4,000 that matched did.

What to Check Before You Trust a Public Dataset

• Camera parity: does the mount height, field of view, and sensor type in the dataset match your actual robot rig, not just a similar-looking one?
• Task granularity: is the labeled action the same task you need, or a broader category that only looks the same in a summary table?
• Environmental range: was the data captured across enough lighting, surface, and clutter variation to represent your deployment site, not just a clean lab?
• Provenance documentation: can you trace who captured each session and under what protocol, the same way a research firm would document interviewer training?
• Annotation source: were labels applied by someone who understood the task, or outsourced at volume with no domain context?

Frequently Asked Questions