Brain-Wave Sensors Could Solve Robotics' Data Bottleneck, Encord and Zander Labs Experiment Shows

Brain Waves as a Training Signal for Robot Manipulation

According to TechCrunch AI, Encord, a data-tooling startup serving AI-model training, is running a trial with Zander Labs, a German neuroscience company, to tag robotic manipulation tasks with electroencephalography (EEG) signals. The experiment involves a human operator wearing a brain-wave-sensing headset while performing a physical task—in this case, extracting wooden blocks from a Jenga tower in Encord’s San Leandro, California warehouse. Rather than treating the human performance as mere demonstration, Encord and Zander are capturing the operator’s neural activity alongside visual and action data, betting that error-detection and intent signals can create a richer training dataset than video alone.

Vineeth Velmurugan, Encord’s head of robot learning and a veteran of OpenAI’s robotics lab and warehouse-automation firm Berkshire Grey, frames this as a response to a fundamental constraint: robotics companies building end-to-end learning systems lack sufficient real-world training data. The shortage is more acute than the text-scarcity problem that language models faced. LLMs trained on internet-scale corpora; by contrast, collecting and annotating physical manipulation data at YouTube-scale is economically infeasible for most vendors.

The Data Scarcity Problem in Physical AI

The robotics industry has known for years that simulation-only training produces models that fail on real hardware. Self-driving car companies sidestep this by collecting their own data, but that approach does not generalize—each company operates in different geographies and weather conditions, limiting reuse. Video-based learning from human demonstrations offers a shortcut but sacrifices sensor fidelity and ground-truth action labeling.

According to TechCrunch AI, Velmurugan estimates that the industry needs a training corpus approximately five times larger than YouTube’s entire video library to achieve the generalization required for broad manipulation tasks. This scale gap explains why Encord has shifted from pure software (data annotation and model evaluation) to becoming a data-producer itself, working with unnamed leading robotics firms to generate datasets in-house.

Why Brain Signals May Unlock Efficiency

Lucas Gehrke, a neuroscientist at Zander Labs supervising the trial, emphasizes that brain-activity patterns during a task indicate cognitive load and decision criticality. Moments where the human operator’s neural signals spike—flagging error recognition or surprise—signal to the model which parts of a sequence warrant the most careful learning. Rather than treating all frames of a video equally, brain-wave-tagged data allows robotics models to focus learning effort on decision points rather than routine movements.

This is not a foundational innovation in neuroscience; it applies existing EEG-inference techniques to a new domain. The novelty is in Encord’s hypothesis that coupling this cognitive metadata to real-world robotic task data will improve training efficiency and model performance—a claim the company intends to test before scaling beyond the trial.

Why This Matters

If Encord’s trial demonstrates that brain-wave-tagged data improves model performance on real tasks, it could accelerate the timeline for generalist manipulation capabilities in both humanoid and warehouse robotics. The constraint shifting from model architecture to data availability means that even incremental improvements in data efficiency compound across the industry. Teams building robotic systems will face a new sourcing decision: whether to invest in neuroscience-grade sensing infrastructure during human demonstrations, or to continue scaling lower-fidelity video capture. The success or failure of this trial may influence funding allocation within robotics startups toward data-infrastructure vendors like Encord rather than toward foundational model research alone.