Brain-Wave Data and the Physical AI Training Bottleneck Holding Back Robotics

The Data Wall That Is Blocking Physical AI From Scaling

Inside a warehouse in San Leandro, California, a worker named Andrew Ceja is carefully pulling wooden blocks from a Jenga tower while wearing a headset fitted with a camera and brainwave sensors. It sounds like a tech-world novelty act, but what Ceja is doing sits at the very frontier of physical AI training data — and it points to one of the most consequential bottlenecks facing the robotics industry today. The company behind this experiment is Encord, a data tooling firm that builds the infrastructure used to train AI models. And the challenge it is trying to solve is not a model architecture problem. It is a raw data problem, and it is severe.

The broader AI industry spent years scaling large language models by vacuuming up text from across the internet — Stack Overflow, Wikipedia, Reddit, the entire written corpus of the web. That strategy worked spectacularly for chatbots and code assistants. But teaching robots to manipulate physical objects in the real world requires something entirely different: dense, high-fidelity recordings of human movement, intent, and error — data that does not exist at scale anywhere on the internet. According to Vineeth Velmurugan, Encord's head of robot learning and a veteran of OpenAI's robot lab and warehouse automation firm Berkshire Grey, it will take a data set roughly five times the size of YouTube's entire video corpus before physical AI models can meaningfully break through their current performance ceiling. "The data simply does not exist," he has said plainly.

This is the core tension now reshaping the robotics investment landscape. Whereas the frontier labs building LLMs could rely on data that cost next to nothing to collect, the teams building physical AI must manufacture their training data from scratch — at enormous expense. That economic reality is turning data generation itself into an industry, not just a research exercise, and it carries significant implications for developers, IT decision-makers, and policy professionals thinking about the infrastructure costs and regulatory landscape of next-generation automation.

Can Brain Waves Actually Improve Robot Learning?

The brain-wave headset Ceja wears was built by Zander Labs, a German neuroscience startup operating on the hypothesis that measuring neural activity during physical tasks can enrich training data in ways that cameras alone cannot. The sensors detect mental states — moments of error, surprise, or heightened intent — that are invisible to conventional recording equipment but highly informative for model builders. Lucas Gehrke, a Zander neuroscientist supervising the Encord collaboration, explains that the amount of brain activity at any given moment during a task offers clues about when a robot model needs to deploy its most computationally intensive reasoning versus when it can coast on lower-effort inference.

Encord describes this as a trial run. The goal is to build an initial brain-wave-tagged data set, run it through customer robotics models, and evaluate whether it produces measurable performance gains before deciding whether to scale the approach. Velmurugan calls it the "bleeding edge" of the effort to solve the robotics data bottleneck — a candid acknowledgement that this is experimental territory, not an established pipeline. That kind of epistemic honesty is notable in an industry prone to overclaiming, and it reflects the genuine uncertainty about which data modalities will ultimately prove most valuable for training physical AI systems.

For developers and engineers building or procuring robotic systems, the Zander collaboration raises a question worth tracking: if brain-state metadata consistently improves model performance on dexterous manipulation tasks, it could eventually become a standard annotation layer in physical AI training pipelines, much as bounding boxes and segmentation masks became standard in computer vision. The regulatory implications — particularly in Europe, where biometric data collection is governed by the GDPR and, increasingly, the EU AI Act — would be substantial. Collecting brain-wave data from human workers in a commercial context is precisely the kind of sensitive processing that regulators are watching closely, as Wired has reported in its coverage of BCI privacy concerns.

Why Egocentric Video and Teleoperation Are Now Core Infrastructure

Beyond brain waves, Encord is building its physical AI training data business around two primary modalities that are rapidly becoming industry standard. The first is egocentric video: footage captured by cameras worn by human workers, often supplemented with additional fixed camera angles and sensor streams. Encord draws this kind of data from several factories around the world, giving it a global footprint for capturing diverse manipulation scenarios. The second is teleoperation data, collected using leader-follower robotic arm rigs where a human operator directly controls one arm while a paired arm mirrors its movements — generating simultaneous recordings of both human intent and robotic execution.

When TechCrunch visited Encord's San Leandro facility, pilots were using these rigs to demonstrate tasks including pouring coffee from a pot into mugs and stacking poker chips — mundane activities that are genuinely difficult for robots to perform reliably. Storage racks in the facility held fake flowers, plastic vegetables, kitty litter trays, bags of wires, and books: the props needed to teach manipulators how to handle household objects. Another pilot, Sofia Infante, was using robotic arms to plug and unplug ethernet cables from servers — a task that data center operators have long wanted automated but that remains out of reach because robotic pincers lack the dexterity and degrees of freedom that human fingers take for granted.

Encord annotates all of its data sets with natural-language descriptions of what each video frame contains — "right hand tightens bolt," for example — to support LLM-based robotic models that need to understand the semantic content of physical actions. Velmurugan estimates that this kind of dense annotation is worth one hundred times as much as raw, unstructured egocentric footage for training specific manipulation skills, and that it costs only twenty times more to produce. On paper, that is a favourable ratio. In practice, it underscores the fundamental cost asymmetry between text-based AI and physical AI: there is no equivalent of scraping the web for robot training data. Every useful data point has to be actively manufactured, which changes the unit economics of the entire field, as researchers at institutions like UC Berkeley's BAIR Lab have documented.

Beyond Video: Muscle Sensors and 3D Hand Reconstruction

Encord is also experimenting with a third data modality that could prove particularly valuable for fine-grained dexterous tasks. A set of sensors strapped to the forearm detects the electrical signals produced by muscles during hand movements. Because egocentric video often fails to capture the full geometry of a human hand — fingers are frequently occluded, lighting is inconsistent, and the camera angle may not show the underside of a grip — Velmurugan wants to use these muscle signals to reconstruct a continuous 3D representation of hand position throughout a task. This would give robot models a more complete and robust understanding of human grasp mechanics than video alone can provide.

This multimodal approach — combining egocentric video, muscle electrical signals, brain-wave metadata, and dense text annotation — reflects a broader industry hypothesis: that no single data stream is sufficient on its own to train robot models capable of generalising across the messy variability of real-world environments. Research published by teams at Stanford and Carnegie Mellon on robot learning from human demonstrations has consistently shown that richer, higher-fidelity data leads to significantly better generalisation, even when the volume of training data is held constant. The challenge is that richer data is exponentially more expensive to produce, which is precisely why companies like Encord are positioning themselves as infrastructure providers rather than end-product robotics firms.

For IT and infrastructure decision-makers evaluating robotic automation procurement, this matters for a practical reason: the quality of the training data your vendor's models were built on will increasingly determine whether those systems can handle edge cases in your specific environment. A robot trained on impoverished, low-annotation data may perform well in demos and poorly in deployment — a pattern already familiar from enterprise software procurement more broadly. As the McKinsey Global Institute's AI research has noted, deployment failures in AI systems are disproportionately traceable to data quality issues rather than model architecture weaknesses.

Comparing Physical AI Data Modalities: What Each Approach Delivers

What makes Encord's position strategically interesting — beyond the novelty of the modalities it is testing — is its cross-industry vantage point. Because it works with many robotics companies simultaneously (Velmurugan says he is not authorised to name them, but describes them as leading firms), Encord can observe which data collection techniques are gaining traction across the industry before any single customer

Originally reported by TechCrunch. Summarised and curated by European Purpose.