Are brain waves the next unlock for physical AI?

Encord Trials Brain-Wave Sensing to Solve Robotics Data Bottleneck

Data infrastructure company Encord is testing a novel approach to a persistent problem in physical AI: the severe lack of high-quality training data. In its San Leandro warehouse, human trainers, or "pilots," are performing complex manipulation tasks like playing Jenga while wearing headsets from German neuroscience startup Zander Labs. These headsets not only capture a first-person view but also measure the pilot's brain waves to deduce mental states like intent, error, and surprise. The initiative represents a move to solve the data bottleneck not by simply managing existing data, but by manufacturing new, higher-fidelity datasets that could significantly accelerate the development of capable humanoid and warehouse robots.

New Modalities for High-Fidelity Data

The collaboration with Zander Labs is part of a broader strategy at Encord to create rich, multimodal data that goes far beyond standard video. According to Vineeth Velmurugan, Encord’s head of robot learning, the goal is to create datasets that are densely annotated and far more valuable for fine-tuning specific skills. The company is experimenting with several data generation techniques to supply what Velmurugan says is a non-existent market for physical training data.

• Brain-Wave Integration: Using Zander Labs' headsets to tag video data with neural signals, potentially helping models identify critical moments in a task.
• Leader-Follower Rigs: Paired robotic arms where a human directly controls one arm and a second one mimics the movement, generating data for tasks like stacking poker chips or pouring coffee.
• Muscle Signal Sensing: Forearm-strapped sensors detect electrical signals in muscles to create a more complete 3D model of hand movements, which are often obscured in video.
• Egocentric Video: Data is also collected from workers wearing cameras in factory environments around the globe.

The Shifting Economics of Physical AI

This pivot towards data manufacturing highlights a fundamental economic difference between training physical AI and large language models. While LLMs were built by scraping massive text corpora from the internet at minimal cost, creating physical data is an expensive, labor-intensive process. Velmurugan estimates that producing densely annotated, high-quality data costs 20 times more than collecting simple egocentric video, but argues it provides 100 times the value for training. This economic reality is turning data generation into its own specialized industry, with companies like Encord positioned to leverage their cross-customer insights to identify the most effective data collection techniques for the entire robotics sector.

The shift from merely managing robotic training data to actively manufacturing high-fidelity, multimodal datasets signals a fundamental change in the AI value chain, where the cost and quality of physical-world data, not just model architecture, will determine the winners in physical AI.