Brain wave robot training has arrived in a warehouse in San Leandro, California, and it looks like a Jenga game. Encord, a data-tooling startup that recently closed a $60 million Series C led by Wellington Management, is experimenting with headsets that measure brain activity while human operators disassemble block towers, pour coffee, and stack poker chips, all in the service of training robot models. The idea: that mental states, not just physical movements, might be the missing signal for physical AI.
The headset in question was built by Zander Labs, a German-Dutch deep-tech company founded in 2016. Its hardware product, the Zypher Suite, is a mobile EEG system designed to monitor brain activity and provide real-time neuroadaptive data with local processing. Lucas Gehrke, a Zander neuroscientist supervising the Encord trial, says the volume of brain activity at any moment in a task reveals when a model needs to deploy its highest-effort processing, essentially flagging where the hard bits are.
Encord’s work with Zander is still a trial run. The plan is to build an initial brain wave-tagged data set, run it through customer robotics models, and evaluate whether it actually moves performance metrics before deciding whether to scale. That is a sensible way to frame a bet on genuinely unproven ground.
Brain Wave Robot Training Gets Serious, But the Data Problem Is Bigger
Vineeth Velmurugan, Encord’s head of robot learning and a veteran of OpenAI’s robot lab and warehouse-automation firm Berkshire Grey, describes brain wave work as the ‘bleeding edge’ of addressing the robotics data bottleneck. The bottleneck itself is less exotic: the physical world simply has not generated enough labelled manipulation data to train capable robot models at scale.
‘The data simply does not exist,’ Velmurugan says. He estimates it will take a data set roughly five times the size of YouTube’s entire video corpus to break through. That figure helps explain why data generation has become a commercial business rather than a side project inside a research lab.
Encord collects two main types of training data. The first is ‘egocentric’ video, captured by workers wearing head-mounted cameras, often augmented with extra camera angles and sensor readings. The second is data from leader-follower rigs, paired robotic arms where a human controls one and the other mimics it. When TechCrunch visited the facility, pilots were using these rigs to practise pouring coffee and stacking poker chips. Storage racks held fake flowers, plastic vegetables, kitty litter trays, bags of cables: the props list for teaching robots to handle household objects.
Encord is also developing a forearm sensor that reads electrical signals in muscles, aiming to reconstruct a 3D model of hand position for moments when camera footage loses sight of the fingers. Each data set gets dense text annotation, ‘right hand tightens bolt’ style, to help language-model-based robot brains parse what is happening. Velmurugan estimates that kind of annotation is worth 100 times as much as raw egocentric footage for training specific tasks, and costs only 20 times more to produce. On paper, a good trade. In practice, 20 times more than near-zero is still a real number.
The Economics That Make Physical AI Different from the LLM Playbook
This is where the comparison between physical AI and large language models runs out of road. LLM builders scraped text from Stack Overflow and the rest of the open web for close to nothing. Physical training data has to be manufactured by people in warehouses with sensors on their forearms and, now, electrodes on their heads. The cost structure is categorically different.
Encord’s own growth suggests the market is real regardless of that friction. According to SiliconANGLE, the data volume on Encord’s platform grew from just over 1 petabyte to more than 5 petabytes in the roughly 18 months before the Series C, and revenue increased more than 10 times over that same period. The Series C, which also drew in existing backers Y Combinator, CRV, N47, and Crane Venture Partners alongside new investors Bright Pixel and Isomer Capital, brings total funding to $110 million.
Zander Labs, for its part, has broader ambitions than robot training experiments. The company signed a contract worth €30 million (approximately $32.9 million) with the German Agency for Innovation in Cybersecurity to develop neurotechnological prototypes under a project called NAFAS, which uses a passive brain-computer interface that does not require users to actively imagine actions. That contract runs until November 2027, according to MassDevice.
Competition is forming. In June 2026, a rival physical-AI data startup called XDOF emerged from stealth with $70 million in funding from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo. Founded in October 2024 by UC Berkeley researchers, XDOF already serves about 20 customers and released a 130,000-trajectory manipulation dataset with Berkeley’s AI research lab, according to Business Model Analyst. The race to become the default supplier of physical training data is, apparently, fully on.
Back in San Leandro, Andrew Ceja is rebuilding the Jenga tower and pulling it apart again. He previously worked at Scale, another AI data-annotation firm, before that at a waste-management company where he kept a robotic trash sorter running. Now he annotates his own brain waves for the benefit of robot models that do not yet exist at scale. ‘It’s something new every day!’ he says. The brain wave data may or may not move the needle on model performance. The question Encord will answer first is whether the signal is real. The economics question comes after that.
