Robots have a reading problem. Not literacy, exactly, but a shortage of the material they need to learn from.
Robotics labs are now collecting motion and video data at scale to train the AI models that power their machines. The goal is simple to state and hard to pull off: teach robots to handle real-world tasks by showing them how humans do it.
The push comes as the industry runs into what amounts to a data bottleneck. Building a robot that can fold laundry or sort parts requires examples of those tasks being performed. Lots of them.
More footage, better robots
The clearest evidence that this approach works comes from Dyna Robotics. Its Dyna-2 model saw task success climb from 20% to 53% in August 2026.
The change behind that jump was the training dataset. Dyna expanded it from 1,000 hours to over 1 million hours of egocentric footage.
Egocentric video is first-person footage, the kind captured from the viewpoint of whoever is doing the work. It shows the hands, the objects, and the sequence of movements from the angle a robot would actually experience.
That result helps explain why massive human video pretraining has become a central pillar of robot development. It matters most for manipulation, meaning grabbing and handling objects, and for locomotion, meaning moving through physical space.
Building the data infrastructure
Innodata opened a motion-capture research and development lab in Ridgefield Park, New Jersey, on September 30, 2026. The facility uses Vicon’s infrared optical cameras to track human and robot movements with sub-millimeter accuracy.
Agile Robots announced its Robot Academy in Munich around October 8, 2026. The academy uses hand-tracking technology to turn ordinary videos of workers into usable trajectory data. That lets the company skip teleoperation rigs entirely.
Companies like Scale AI are operating data factories and robot gyms across multiple countries, including Mexico, China, and Indonesia. Recording setups are distributed globally, and people perform tasks that get captured and fed into training pipelines.
The money behind the movement
Robotics and physical AI companies have raised nearly $48 billion in funding so far in 2026.
That capital is landing during a notable shift in strategy. Developers are moving away from datasets generated mainly through simulation and toward large collections of human video and precision motion data.
A generalist policy is a model that can handle many different tasks rather than one narrow job. It’s the difference between a factory arm that welds one seam forever and a machine that could plausibly be asked to do something new tomorrow.
What this means for the robotics race
The three approaches on display represent different tradeoffs. Innodata’s Vicon-equipped lab offers precision but limited throughput. Agile Robots’ worker-video method promises scale with less hardware. Distributed robot gyms trade some control for volume and geographic diversity.
There are open questions worth watching. Distributing data collection across countries like Mexico, China, and Indonesia raises issues around labor conditions, consent, and how footage of workers is used. Turning ordinary job videos into training material, as Agile Robots plans, will likely invite similar scrutiny.
Dyna went from 20% to 53% with a thousandfold data increase. A 53% success rate still means nearly half of attempts fail, which is not yet a number most factories or homes would accept.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.