How watching humans work could teach the next generation of factory robots
A robotics engineer explains how first-person camera footage of people doing real tasks could cut the time and cost of training industrial robots.
Key points
- Ted Larson, CEO of OLogic, will present on egocentric robot learning at RoboBusiness on 21 October 2026 in Santa Clara, California.
- Egocentric learning trains robots using first-person video of humans doing real tasks, rather than hand-coded rules or remote control.
- NVIDIA, Google DeepMind, and Meta are all investing in versions of this approach.
- Industries from manufacturing to healthcare stand to benefit if the technique can cut deployment time and reduce the need for specialist engineers.
Imagine strapping a camera to your head while you pack boxes or help a patient sit up in bed. A robot watches that footage and learns to do the same job. That's the basic idea behind egocentric robot learning, a training method that feeds robots first-person human video instead of programming every move by hand.
The approach is attracting serious attention. NVIDIA, Google DeepMind, and Meta are all working on it, alongside a growing number of specialist startups building what researchers call foundation models for robotics: broad, flexible AI systems trained on large amounts of movement data, similar to how large language models, the technology behind chatbots like ChatGPT and Claude, are trained on text.
Why does this matter for ordinary workplaces?
Deploying a robot in a warehouse or a hospital ward typically takes months of specialist engineering. Egocentric learning promises to shorten that by letting robots borrow skill directly from the people already doing the job.
Labour shortages are real across manufacturing and healthcare. If a robot can be trained by recording a worker for a few hours rather than programming it for months, smaller organisations that can't afford bespoke automation get a practical option for the first time. Our coverage of Vention's Physical AI Lab in Montreal, where robots learn from live production data rather than controlled experiments, shows this pressure is already reshaping how companies approach training from scratch.
That's the promise. The proof is still partial. Most results come from controlled lab settings, and scaling these systems to messy, unpredictable real-world environments is genuinely hard. Imitation learning, teaching a machine by showing it examples rather than writing explicit rules, has a long history of working well in the lab and struggling outside it.
What will the RoboBusiness session cover?
Ted Larson, CEO and co-founder of OLogic, an embedded systems research and development firm that has contributed engineering work to robots for Bear Robotics, Simbe Robotics, and others, will lead the session at 10:45 a.m. On the second day of RoboBusiness 2026. The Robot Report covered the announcement.
Larson holds computer science degrees from California Polytechnic State University and has more than 25 years in robot and consumer electronics design. He plans to cover the foundations of the technique alongside its practical limits, including multimodal data collection, meaning gathering video, audio, touch, and motion data together, and how large-scale training datasets are changing what commercial robots can do.
The honest read: this is a conference session promoting an emerging technique, not a peer-reviewed result. The underlying research thread is real, the industry investment is real, and the workforce problem it aims to solve is real. The gap between promising demos and reliable deployment is where this field will be won or lost.
Common questions
Does this mean robots are recording workers without their knowledge?
Not inherently. Egocentric learning systems use footage collected deliberately for training, with workers wearing cameras as part of a structured demonstration process. How that data is collected, stored, and consented to is a workplace policy question every employer adopting these tools will need to answer.
Is this technology ready to use today?
Some narrow versions are in commercial use, for example in warehouse picking systems. Broader, flexible versions that can learn varied tasks from short demonstrations are still largely in research and early-deployment stages as of mid-2026.
