In short
Encord is experimenting with brain-wave and muscle-sensor data to improve robot training in San Leandro. The project reflects a broader push to solve physical AI’s biggest problem: a shortage of high-quality real-world data.
- Encord is testing brain-wave and muscle-sensor inputs to see whether they improve robot training data.
- The company believes physical AI is constrained less by model design than by scarce real-world training examples.
- Dense annotation and teleoperated demonstrations may be more valuable than simple video for robotics.
- The San Leandro warehouse functions as a data factory for tasks that humanoid and warehouse robots need to learn.
Encord is testing whether brain-wave signals can improve robot training, and the experiment could shape how humanoid and warehouse robots are built in the years ahead. At a San Leandro, California, facility, the data company is combining video, motion tracking and neuroscience tools to create rare training datasets for physical AI.
The work matters because robotics companies are running into a fundamental bottleneck: there is far less useful training data for machines that move and manipulate objects than there is for language models trained on the internet. Encord’s approach suggests that the next breakthrough in physical AI may depend less on model design and more on manufacturing better data.
On one side of the room, a pilot carefully dismantles a Jenga tower while wearing a headset that records what he sees. On another, robotic arms practice tasks such as pouring coffee, plugging cables into server racks and stacking small objects. The goal is not just to observe these actions, but to capture signals that may reveal intent, surprise and error — including brain activity itself.
Why Encord is experimenting with brain waves
Encord’s central idea is straightforward: if robot builders can better understand what a human operator is thinking during a task, they may be able to train machines more efficiently. The company is testing that theory with Zander Labs, a German neuroscience startup whose hardware measures brain activity during physical work.
Rather than trying to prove the concept in theory, Encord is running a live trial. The company wants to gather an initial dataset tagged with brain-wave information, feed it into robotics models and see whether performance improves. Only if the results are promising would it make sense to expand the program.
That cautious approach reflects how experimental the field still is. Brain signals are not yet a standard robotics input, and no one is suggesting they will replace video or motion capture. Instead, the hope is that they can add another layer of context, especially in situations where a human operator may be deciding how much effort to apply or when a task has gone wrong.
Encord’s team describes the brain-wave work as an early test designed to determine whether this kind of signal can make robot training data more useful before the company scales it further.
How the San Leandro robot data lab works
The San Leandro warehouse is less a traditional robotics lab than a production line for training examples. Encord employs pilots — its internal term for the people generating data — who perform tasks while cameras and other sensors record the process from different angles.
Some of the setups use leader-follower robotic arms. In those rigs, a human directly controls one arm while a second arm mirrors the movement, creating paired examples that are useful for training manipulation models. The company also stages task-specific scenarios, from household chores to industrial procedures.
When TechCrunch visited the site, the tasks included:
- pouring coffee into mugs
- stacking poker chips
- plugging and unplugging ethernet cables from server equipment
- handling fake flowers, books, plastic produce and other household objects
These may look simple, but they expose exactly where robot dexterity still falls short. Grippers can grab and move things, but they cannot yet match the flexibility of a human hand. Tasks that appear mundane to people become highly complex when a machine must execute them repeatedly, precisely and safely.
What the physical setup reveals about robotics today
The San Leandro demonstrations show that physical AI is still constrained by the limitations of hardware. Even when software understands the task, the robot may lack the fine motor control needed to carry it out reliably.
That is why Encord’s facility is stocked with an eclectic set of props: wires, vases, books, small containers, artificial fruit and other everyday items. The company is trying to create enough variation to teach models how objects feel, move and respond in the real world.
One of the most telling tasks is the server-cable exercise. Data-center work is a natural target for automation, but the precision required to insert and remove connectors remains difficult. The demo illustrates a broader truth about robotics: an action that seems trivial to a person can still be out of reach for a robot.
What makes physical AI different from language models?
Physical AI faces a harsher data problem than large language models do. Language systems were trained on enormous volumes of text scraped from across the internet, giving them access to a vast and relatively cheap supply of examples. Robots do not have that luxury.
Robotic systems need data that reflects physical interaction: touch, force, timing, spatial positioning and the consequences of mistakes. Some of that can be captured from video, but video alone often fails to preserve enough detail about what the operator was actually doing or intending.
Encord’s head of robot learning, Vineeth Velmurugan, said the field is now hitting a structural wall. In his view, the relevant data simply does not exist in the quantities modern robotics companies need, so the data must be created from scratch.
That changes the economics dramatically. Instead of pulling free text from the web, robot teams must pay for labor, equipment, sensors, staging, storage and annotation. The result is a business model built around data manufacturing rather than data collection.
| Aspect | Language AI | Physical AI / Robotics |
|---|---|---|
| Primary training source | Internet text and code | Human demonstrations, robot teleoperation, sensor data |
| Data availability | Massive and widely available | Limited and must often be created |
| Cost structure | Relatively low marginal cost | High labor and equipment costs |
| Common challenge | Reasoning and hallucination | Dexterity, precision and real-world variability |
Why dense annotation matters
Encord is not only collecting raw demonstrations. It is also attaching detailed descriptions to each action so models can understand what is happening in context. A note such as “right hand tightens bolt” carries more usable information than a plain video clip on its own.
Velmurugan argues that richly labeled data can be far more valuable than generic egocentric footage because it helps models map motion to intent. The trade-off is cost: the company estimates this denser style of annotation may be around 20 times more expensive to produce than basic footage, even if it could be far more useful for training.
That comparison matters because robotics companies are not simply searching for more data; they are searching for the right data. In some cases, better annotations can be worth the added expense if they help a model learn a task faster or with fewer failures.
Encord’s pitch is that it sits at the intersection of many robotics programs, giving it visibility into which data strategies are actually working. That vantage point could help the company spot winning techniques before a single customer fully commits to them.
How do companies collect robot training data?
They mostly use two approaches: egocentric video from human workers and teleoperated robots that mirror human movement. Both methods are designed to capture the physical steps that a machine must eventually reproduce.
Egocentric data comes from workers wearing head-mounted cameras or similar devices that record what they see as they perform a task. Teleoperation involves a human controlling a robot remotely, often through paired arms or other mechanical proxies.
Encord is using both. It gathers footage from factories in multiple locations, then uses its California site to run experiments with additional sensing methods and task-specific collections. The brain-wave headset from Zander Labs is one such experiment. Forearm sensors that detect muscle signals are another.
These methods are meant to solve a common problem: cameras often miss the full complexity of hand motion. If a hand is partly out of frame or obscured, the training data loses fidelity. Sensors on the arm may help reconstruct the three-dimensional posture of the hand more accurately.
What role do muscle sensors play?
Muscle sensors are intended to fill in details that video alone cannot capture. By measuring electrical activity in the forearm, Encord hopes to infer hand position and movement more reliably, which could make robot training data more complete.
That matters for tasks requiring precise finger placement, grasping and tool use. If a model can better estimate how a human hand is configured in space, it may improve the robot’s ability to reproduce the same action later.
Who is building the workforce behind physical AI?
The workers creating robotics datasets are becoming a specialized labor force in their own right. Encord’s San Leandro operation relies on a small group of pilots who generate demonstrations, operate robotic rigs and help annotate the results.
Two of them, Andrew Ceja and Sofia Infante, previously worked at Scale, another data-labeling company in the AI ecosystem, before joining Encord. Their roles show how the data infrastructure around AI is evolving into a career path, not just a stopgap job.
Ceja’s background is especially telling. He previously worked at a waste management company, where he helped keep a robotic trash sorter running. That experience made him comfortable with machines in messy, practical settings — exactly the kind of environment physical AI must eventually handle.
For these pilots, the work is hands-on and iterative. If a demonstration fails to teach the robot something useful, the team adjusts the task and tries again. The process resembles laboratory research, but with warehouse equipment, props and production goals.
Ceja said he enjoys the challenge of creating training tasks for machines because the work changes constantly and keeps him solving new problems every day.
What Encord’s robotics business says about the market
Encord was originally founded to help companies annotate computer-vision data and evaluate machine-learning models. As customers moved toward end-to-end robotic learning, the company saw an opportunity to expand from data management into data creation.
That shift is important because it suggests the value in robotics may be moving upstream. If the core limitation is not the model but the dataset, then the companies that can produce specialized, high-quality training data may gain significant leverage.
Velmurugan said Encord works with many major robotics companies, though he did not name them. The firm’s position between customers also gives it a broad view of what the industry is asking for: more realistic demonstrations, more precise labeling and more useful signals from the human body.
The company’s broader thesis is that robotics will not advance on model innovation alone. It will also require a data supply chain that can scale, standardize and reproduce useful examples of physical behavior.
What are the limits of the brain-wave approach?
The biggest limitation is that brain-wave data is still experimental and unproven in this context. The signals may turn out to be too noisy, too difficult to interpret or not informative enough to justify the added complexity.
There is also a broader practical issue: collecting this kind of data is expensive and labor-intensive. Even if it improves model performance, the improvement has to outweigh the cost of gathering it.
That is why Encord is treating the project as a trial rather than a full product launch. The company is testing whether the signal is worth the investment before it commits to broader deployment.
Still, the experiment highlights where the field is headed. Robotics companies are no longer just asking how to build larger models; they are asking how to capture richer human behavior in the first place.
A timeline of the physical AI data race
The push into physical AI data has evolved quickly as robotics companies race to create systems that can generalize beyond narrow tasks.
| Period | Development | Why it matters |
|---|---|---|
| Early AI era | Models trained on internet-scale text and code | Showed that massive data could drive rapid capability gains |
| Robotics phase one | Companies relied on limited demonstrations and simulation | Useful, but not enough for highly dexterous manipulation |
| Current phase | Human teleoperation, egocentric video and richer sensor data | Attempts to capture the full complexity of physical tasks |
| Emerging phase | Brain waves, muscle sensors and dense annotations | Could provide deeper context about intent and effort |
Why this matters beyond one warehouse
Encord’s experiment is part of a much larger effort to make robots more capable in the real world. Warehouse automation, data-center maintenance, household assistance and humanoid labor all depend on the same breakthrough: robots must learn from richer, more realistic examples of human movement.
If brain-wave signals or muscle sensors prove useful, they could become part of a new standard for robotics data collection. If they do not, the industry will still have learned something valuable about where the real bottlenecks lie.
Either way, the San Leandro warehouse underscores a key point in the physical AI race. The next leap may not come from a bigger model alone, but from a better understanding of the human body at work — down to the neural and muscular signals that accompany a simple task like pulling a block from a Jenga tower.
What comes next for Encord
For now, the company is continuing to test the limits of its data-generation model. The immediate question is whether brain-wave-tagged examples improve robotics performance enough to justify further investment.
If they do, Encord could expand its use of neuroscience tools and richer sensor packages. If not, the company still stands to benefit from its broader position as a supplier of specialized robotics data.
The more immediate lesson is that physical AI is turning into a data business as much as a hardware or software business. Whoever solves the training-data problem may help decide how quickly humanoid and warehouse robots move from promise to practical deployment.
Frequently asked questions
What is Encord testing in physical AI?
Encord is testing whether brain-wave data, combined with video and motion capture, can create better training datasets for robots. The company is evaluating whether those signals improve model performance before deciding whether to expand the project.
Why are robotics companies struggling with data?
Robotics companies are struggling because physical tasks require much more specific and harder-to-collect data than language models need. Robots must learn force, timing, dexterity and context, and that information usually has to be generated manually rather than scraped from the web.
How does Encord collect robot training data?
Encord uses human operators, teleoperated robotic arms, egocentric video, detailed annotations and additional sensors. In its San Leandro facility, pilots perform tasks such as pouring coffee, stacking objects and handling cables while the systems record the movements.
Why are brain waves interesting for robotics?
Brain waves may reveal intent, effort, error and surprise in ways video cannot. If those signals help builders understand when a human operator is struggling or making a decision, they could make robotics datasets more informative and easier to train on.
