The data that can't be scraped

Large language models became capable because there was a vast amount of text to learn from. Books, code, forums, articles: decades of human writing, already digitised and waiting. Robots don’t have an equivalent. And for the skill we care most about, handling objects that bend by feel, there is almost nothing to start from at all.

The size of the gap

Language models now train on somewhere between 15 and 36 trillion tokens. The largest robot datasets hold about 10,000 hours of camera and joint data. They have no standard touch channel.

Ken Goldberg has called robotics’ shortfall a “100,000-year data gap.”¹ The phrase is useful because it points at the real problem. It is not that robotics researchers are slow. It is that the data robots need does not exist in a form anyone can download.

For touch, the gap is wider still. Frontier robot models learn mostly from camera data. Open manipulation datasets carry almost no tactile signal. Even the robot data that does exist mostly shows what a task looked like, not what it felt like.

Why touch can’t be scraped

Video of people doing things is everywhere. Touch data is not, for a simple reason: nobody records it. When someone routes a cable or packs a soft pouch, the information that decided success, the tension, the slip, the click, leaves no trace.

That information is rich. The human hand has about 17,000 mechanoreceptive units.² A camera filming that hand captures none of their output.

So touch data has to be collected deliberately, with sensors, on real tasks. It must be collected on real floors, one shift at a time.

Force/torque is not a shortcut

It is tempting to think a wrist force/torque sensor solves this. Many robots have one. But force/torque isn’t touch. It measures how hard the arm is pressing in total. It doesn’t measure where contact is on the fingers or whether the object is starting to slip. For deformables, those are the signals that matter.

Why floor time compounds

The industry is starting to agree that touch matters. Figure, for example, moved its Helix policy from vision-only to fingertip touch.³ We expect more teams to follow.

Adding a sensor is the easy part.

Frontier labs can add a touch sensor. They can’t skip the floor time it takes to master deformables.

Skill with deformables comes from contact with real objects in real conditions: cables that are stiffer than spec, bags that stick, pouches filled unevenly. Each of those cases has to be felt, recorded and learned from. Simulation helps with some of it. It does not yet capture how a real poly bag behaves on a real line.

That is why floor time compounds. We think about it as a loop:

1. Capture. Record how skilled hands and robots handle deformables, with touch as a first-class signal.
2. Learn. Train models that use those touch signals to decide how to grip, adjust and confirm.
3. Deploy. Put those models on robots doing real work.

Every deployment adds contact data that makes the next robot better. The loop gets faster the longer it runs.

A channel that can’t be distilled

There is a second reason this data matters. Models get smaller and cheaper over time. A capability that needs a very large model today can often be distilled into a smaller one later. That is good for everyone.

But distillation only works on what a model already knows. If no model was ever trained on touch, there is nothing to distil.

A large model can be distilled; a missing sensory channel can’t.

How we are approaching it

We are building each part of the loop.

Capture. Tacto is a tactile capture system that records how skilled hands handle deformables. Our robots are also a source of contact data as they work.

Learn. Our VLSA model (vision, language, sensory, action) is designed around touch from the start. It pairs a vision-language model of about 3 billion parameters with a Cerebel touch encoder: a roughly 50-million-parameter vision transformer that turns tactile signals into touch tokens. An action expert based on flow matching turns those inputs into motion, closed loop at 100 Hz on the robot.

Deploy. AdiBot is our mobile manipulator, designed and assembled in Dallas and built lean enough to deploy as a fleet. Version 1.0 is built, with a 7-DoF arm, lift, gripper, depth cameras and teleoperation. The next generation is in development, designed with a dual-arm torso, dexterous hands with 48 tactile channels per finger at 100 Hz, and Jetson Thor on board. Our Auto-Executor has been demonstrated on tote-to-tote picking.

What this means

For an operations leader, this means the robots that handle your soft goods well will be the ones that have already handled a lot of soft goods. Early floor time is how that experience is earned.

For a robotics engineer, it means the next useful dataset is not more video. It is synchronised touch, vision and action from real tasks.

For an investor, it means the defensible asset in this part of robotics is not only the model. It is the contact data behind it, and the loop that keeps adding to it.

Deformables are the start. The same touch intelligence transfers to any robot hand. But it has to be learned first, and that learning happens on the floor.