Robotics Firms Accelerated Data Collection for Physical AI
Companies are scaling demonstration datasets to improve robot movement precision and manipulation capabilities.
Updated on Oct. 13, 2026 in Robotics
Live Poll
Do you believe the rapid expansion of AI data collection will benefit society in the future?
Robotics developers are rapidly increasing the volume of training data to improve autonomous performance. Researchers now use millions of hours of human video and high-fidelity motion capture to train new physical AI models.
Why it matters
Improving robot performance in granular manipulation and movement tasks requires massive, diverse datasets. This shift toward large-scale data collection attempts to bridge the gap between static programming and generalized physical intelligence.
Dyna Robotics achieved a 53% normalized score across 14 tasks using the Dyna-2 model, which was pretrained on 1 million hours of human video.
The players
Dyna Robotics
A developer of physical AI models focused on training architectures using large-scale human video datasets.
Innodata
A data engineering company that operates motion-capture research facilities to provide high-fidelity training data for robotics.
Scale AI
A provider of data infrastructure that processes thousands of hours of demonstration data daily to train autonomous systems.
Universal Robots
A manufacturer of industrial collaborative robots with over 100,000 active deployments globally.
The details
The industry is increasingly relying on motion capture to translate human physical nuances into robot control signals. Innodata’s new lab in New Jersey utilizes infrared optical cameras to record three-dimensional movement, while Scale AI aggregates over 1,000 hours of demonstration data daily from residential and commercial deployments to train models like those used by Universal Robots.
Timeline
March 16, 2026: Scale AI entered into a partnership with Universal Robots.
Sept. 30, 2026: Innodata launched its motion-capture research facility in New Jersey.
The Tech Race
The transition toward large-scale data pretraining represents a clear departure from traditional manual robotic programming. This industry-wide effort now mirrors the high-parameter model race seen in natural language processing, but applied to physical interaction.
The immediate benefits of these data-driven improvements will appear in industrial settings where complex manipulation is required. Expect these performance gains to gradually filter into commercial and residential robot assistants as models achieve higher success rates in unstructured environments.
The takeaway
The race to achieve generalized robot capability is now defined by the sheer volume of human demonstration data used for training. Watch for upcoming performance benchmarks from major manufacturers as they integrate these 1-million-hour-scale models into their standard hardware deployments.
Further reading
Learn more about the latest advancements in autonomous systems within the Robotics section.
Source note: This article includes information reported by TokenPost.
Live Poll
Do you believe the rapid expansion of AI data collection will benefit society in the future?
