Encord and Zander Labs trial brain wave data for robot training

New Approach to Robotics Training Data

Encord, a company focused on data tooling for AI models, is exploring a novel method for training robotic systems. They are collaborating with Zander Labs, a German neuroscience startup, to integrate brain wave data into their training datasets. This trial involves human trainers wearing headsets that capture both visual information and brain activity while performing tasks like disassembling a Jenga tower.

Addressing Data Scarcity in Robotics

The initiative stems from the belief that the primary limitation for advanced robotics, particularly humanoid and warehouse robots, is the lack of sufficient real-world physical training data. Encord is moving beyond simply managing existing data to actively manufacturing new, more comprehensive datasets. Zander Labs' technology aims to deduce mental states such as error, intent, and surprise from brain activity, which could provide richer context for AI models.

Trial and Evaluation

The current collaboration is a trial run to create an initial dataset tagged with brain wave information. This data will then be used to train customer robotics models, and its effectiveness in improving performance will be evaluated before any decision to scale up the approach. Lucas Gehrke, a neuroscientist at Zander, suggests that brain activity levels during tasks could help model builders determine when to deploy more intensive models.

Encord's Role in Robotics Data

Encord was originally founded to assist companies with annotating data and evaluating models for machine-vision applications. As their robotics clients began adopting end-to-end learning for manipulation tasks, Encord recognized the need to generate training data directly, rather than just managing it, due to its scarcity. Vineeth Velmurugan, Encord's head of robot learning, describes this effort as the "bleeding edge" in solving the robotics data bottleneck.