Encord, a San Leandro, California firm that supplies data‑annotation tools for machine‑vision applications, has begun a trial that pairs human brain‑wave recordings with robot‑training video. The experiment, conducted in the company’s warehouse‑turned‑lab, has pilot Andrew Ceja wearing a headset built by German neuroscience startup Zander Labs while he carefully removes blocks from a wobbling Jenga tower. The headset captures electrical activity in Ceja’s brain, adding a new layer of information to the visual footage that Encord already collects for its customers.

Encord’s head of robot learning, Vineeth Velmurugan, says the goal is to create an initial brain‑wave‑tagged dataset, run it through client robotics models, and measure any performance lift before deciding whether to scale the approach. Zander Labs neuroscientist Lucas Gehrke explains that the intensity and timing of neural signals can hint at moments of error, intent or surprise, offering model builders clues about when to deploy higher‑effort algorithms.

The trial reflects a growing consensus among robotics companies that existing data sources are insufficient. Velmurugan, a veteran of OpenAI’s robot lab and Berkshire Grey, notes that customers often lack the physical‑world footage needed to train end‑to‑end learning systems. “The data simply does not exist,” he says. Encord therefore positions itself not just as a data‑management platform but as a data‑manufacturing business, gathering egocentric video from workers in factories worldwide and supplementing it with novel modalities like brain‑wave and forearm‑muscle sensors.

In the San Leandro facility, pilots use leader‑follower rigs—paired robotic arms where one mimics a human operator’s movements—to capture tasks ranging from pouring coffee into mugs to stacking poker chips. Another station has Sofia Infante maneuvering a robot arm to plug and unplug Ethernet cables from a server rack, a chore data‑center operators would love to automate. A set of sensors strapped to the forearm records electrical muscle activity, which Velmurugan hopes will enable a 3‑D reconstruction of hand position, filling gaps that standard egocentric video often misses.

Encord’s datasets are densely annotated with physical descriptions such as “right hand tightens bolt.” Velmurugan estimates that such granular labeling is worth about 100 times the value of generic ego video, even though it costs roughly 20 times more to produce. The trade‑off underscores a key difference between building large‑language models, which can scrape massive text corpora at minimal cost, and training physical‑AI systems that require bespoke, high‑fidelity data.

While the brain‑wave trial remains in its early stages, the company believes the added signal could help robot developers pinpoint moments when a model needs to apply its most sophisticated reasoning. If successful, the approach could narrow the gap between the massive video datasets that fuel vision models and the precise, task‑specific data robots need to manipulate objects with human‑like dexterity.

Encord’s broader strategy leverages its visibility across the robotics ecosystem. By working with dozens of pilots and feeding data back to multiple customers, the firm can spot emerging techniques before any single client adopts them. That insight, Velmurugan says, keeps the dozen or so pilots at the San Leandro site busy and positions Encord as a central hub in the nascent physical‑AI supply chain.

Both Ceja and Infante previously worked at Scale, another AI data‑annotation firm, before joining Encord. Ceja’s background includes stints at a waste‑management company where he maintained a robotic trash sorter. He describes the work as “something new every day,” reflecting the experimental nature of building the data foundations for tomorrow’s robots.

Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.