Robot training data is moving beyond camera footage into brain-wave signals, TechCrunch reported, with Encord testing a Zander Labs headset at its San Leandro facility to see whether human intent, error and surprise can improve physical AI models.

Brain-Wave Trial Tests A Robotics Data Gap

The work is still a trial rather than a production data product.

Encord’s plan is to build an initial brain-wave-tagged dataset, run it through customer robotics models and decide whether performance gains justify scaling the method.

Zander Labs built the headset used in the trial.

The German neuroscience startup’s sensors are designed to infer mental states while a human trainer performs physical tasks, adding a signal that ordinary video does not capture.

Encord Manufactures Data Rather Than Scraping It

Encord began as a company that helps machine-vision teams annotate data and evaluate models.

Its robotics work now extends into producing training material because customers applying end-to-end learning to manipulation tasks need physical-world datasets that are not already available online.

Vineeth Velmurugan, Encord’s head of robot learning, said the needed dataset could be about five times the size of YouTube’s video corpus.

That scale explains why physical AI data collection is becoming an operating business instead of only a research function.

San Leandro Facility Adds New Modalities

The San Leandro site combines egocentric video from workers wearing cameras, leader-follower robotic arms and data gathered from several factories.

TechCrunch visited stations where trainers created examples for pouring coffee, stacking poker chips and handling objects used in household-task datasets.

One task involved plugging and unplugging ethernet cables from the back of a server, work that data-centre operators may want automated but that still requires fine manipulation.

The limitation is physical: robotic pincers remain less dextrous than human fingers and do not match human arm movement.

Encord is also developing forearm sensors that detect electrical signals in muscles.

The goal is to reconstruct a fuller three-dimensional view of hand movement when camera footage misses parts of the hand.

The facility shows how physical AI data becomes a labour and operations problem.

Racks at the site held flowers, books, plastic vegetables, kitty litter trays, scoops and wiring bundles for household-task collection.

About a dozen pilots produce building blocks for neural networks, and several trainers previously worked at Scale, another AI data annotation company, before moving into robotics data work.

Dense Annotation Raises The Cost Of Physical AI

The company annotates physical datasets with descriptions such as a right hand tightening a bolt so language-model-based systems can interpret the action.

Velmurugan estimated that dense annotation can be worth 100 times as much as weaker egocentric data for specific tasks while costing 20 times more to produce.

That cost separates robotics data from the text-heavy history of generative AI.

Internet-scale text could be scraped cheaply; physical data has to be staged, captured, labelled and checked before it can train models for manipulation.

The commercial question for Encord is not whether more signals can be collected.

Customer robotics models still have to show that brain-wave tags, muscle sensors and dense annotation improve task performance enough to pay for manufactured data.