New Step for Physical AI: How Brain Waves Train Robots

In a warehouse in San Leandro, California, a new era in AI is being tested using a game of Jenga. Encord, a company that creates datasets for training AI models, has set out to solve the biggest challenge in robotics: the scarcity of real-world physical training data. An employee of the company, wearing a headset equipped with a special camera and sensors, carefully moves wooden blocks. Techcrunch.com reports this.

According to ixbt.com, this headset not only monitors the environment but also measures the operator's brain waves. Developed by German neurotechnology startup Zander Labs, the technology allows for the formation of more useful datasets to train models by detecting mental states such as error, intention, and surprise. Encord and Zander Labs are currently testing this approach.

Data Shortage for AI

Robotics experts note that the main constraint today is not model architecture, but a severe lack of real physical training data. While chatbots like ChatGPT were trained on internet text, no such ready-made sources exist to teach neural networks physical movements. Collecting data for cars is difficult and expensive, and using videos cannot ensure real-world accuracy.

Vineeth Velmurugan, head of robot training at Encord, emphasizes that the problem of text scarcity is also relevant in robotics. According to him, model creators need a dataset five times larger than the YouTube video corpus. That is why not just managing data, but generating it has become a distinct business direction.

Opportunities Identified Through Brain Waves

Zander Labs neuroscientist Lukas Gerken explains that the volume of brain activity expended during a specific task shows model creators at which stage to deploy the highest-power models. In this testing process, operators remotely controlling robots enrich the database by performing complex tasks such as pouring liquid from a coffee pot into a cup and stacking poker chips.

Currently, Encord is collecting egocentric videos from various factories globally while experimenting with new methods like brain waves at its San Leandro center. If these tests prove successful, the process of training humanoid and warehouse robots is expected to rise to a qualitatively new level in the future.