Brain Waves: A New Weapon for Training AI Robots

In a warehouse in San Leandro, California, an employee sits wearing a helmet equipped with a camera and neural sensors, carefully pulling wooden blocks from a wobbling tower in full concentration. The scene looks ordinary at first glance, but it is in fact a next-generation experiment: measuring human brain waves with the goal of improving robots' ability to learn and understand the physical world around them.

The Data Crisis: The Real Obstacle Facing Tomorrow's Robots

The physical AI industry — the field concerned with building robotic systems capable of interacting with their surrounding environment — faces a fundamental challenge that has nothing to do with model design or processing power. The real bottleneck lies in the scarcity of real-world training data. Unlike large language models, which built their knowledge base from billions of web pages at minimal cost, robots require precise physical data that is difficult to collect or replicate digitally.

UK-based Encord, a company specializing in data tools for training AI models, is tackling this crisis with a different approach: rather than simply managing existing data, it now manufactures data from scratch. Vineet Velmoorgan, Head of Robotic Machine Learning at Encord and a veteran of OpenAI's robotics lab, puts it plainly: "The data simply doesn't exist."

Brain Waves: Data from a New Layer

Encord is collaborating with German neuroscience firm Zander Labs to develop a helmet that measures the brain's electrical activity while human operators perform manual tasks. The idea goes beyond recording movements — it extends to capturing mental states such as error, surprise, and intent, converting them into an additional data layer that supplements traditional video footage.

Lucas Gehrke, the neuroscientist overseeing the project, explains that brain activity levels at specific moments give models valuable cues about when to apply the highest levels of computational attention — which translates in practice to more efficient models that waste fewer resources. The experiment is still in its early stages; Encord plans to build a dataset annotated with brain wave data, then test its impact on the performance of actual robotic models before deciding whether to scale up.

Multiple Approaches to Collecting Real-World Data

Encord's pipeline relies on two primary sources for generating training data:

• Egocentric Video: Footage captured from the human operator's perspective via head-mounted cameras, supplemented by additional angles and complementary measurements, collected from factories at multiple locations around the world.
• Leader-Follower Rigs: Paired robotic arms in which a human directly controls one arm while the other precisely mirrors its movements, used to capture data on fine-grained tasks such as pouring coffee and connecting cables.

The team is also experimenting with a third emerging technique: forearm-mounted sensors that read electrical signals in the muscles, with the aim of building a full three-dimensional model of hand movement — something conventional video footage cannot achieve with sufficient accuracy.

Cost: The Decisive Difference Between Robots and Chatbots

Velmoorgan estimates that richly annotated data — such as "right hand tightening the screw" — is up to one hundred times more valuable for training than raw data, while costing only about twenty times more to produce. On paper, that's a winning equation.

Yet "twenty times more" still means real money. This is the fundamental difference between physical AI and its language counterpart: language models built their knowledge wealth from mostly free internet data, while robot data must be generated hand by hand, trial by trial — radically changing the economics underlying the development of these models. Velmoorgan believes that reaching a true breakthrough requires a volume of data roughly equivalent to five times all the video content on YouTube.

Strategic Position: Working with Everyone at Once

Encord's work with a broad spectrum of robotics companies — whose names Velmoorgan declines to reveal — affords the company a rare strategic vantage point on progress across the sector. It can observe which approaches prove most effective and which fall short, then leverage that accumulated knowledge for the benefit of all its clients. In a market where data is the rarest gem of all, this intermediary position amounts to a formidable competitive advantage.

What is happening in the San Leandro warehouse is not merely data collection; it is an emerging model for an entire industry: the real-world data industry, which may prove to be the hidden engine driving the physical AI revolution in the years ahead.

✦ بقلم فريق دروب أيديا