Everything That Bends: why touch is the missing layer of physical AI

Pick up a phone charger cable. Coil it. Plug one end into a laptop. You did all of that without looking hard at anything. Your fingers knew when the cable slid, when it snagged, and when the plug went home. That knowledge is the reason we started Cerebel Robotics. We build tactile intelligence for robot hands: touch-aware AI for everything that bends.

Robots got good at the wrong half

Physical AI has moved fast. Robot models can now look at a scene, name the objects in it, plan a sequence of steps and generate motion to carry it out. Ten years ago that would have sounded like the finish line.

It isn’t. The work that remains is the part humans find too easy to notice. Hans Moravec pointed this out decades ago: the skills that feel effortless to us, like perception and movement, are the hardest to give a machine.¹ Reasoning turned out to be the easier half. Handling turned out to be the harder one.

Nowhere is that clearer than with objects that bend.

What we mean by “everything that bends”

Engineers call these deformable objects. We sort them into three groups:

• 1D: wires, cables, hoses.
• 2D: bags, film, fabric.
• 3D: foam, pouches, soft goods.

They have three things in common. They change shape when you touch them. They hide their own state, so a camera can’t tell you whether a cable is under tension or whether you’re holding one sheet of film or two. And they remain an open research problem. A rigid box behaves the same way every time you grasp it. A poly bag does not.

These objects are not edge cases. They are cable harnesses, shipping bags, garment folds, foam inserts in a kit. They show up in every warehouse, factory and server room. They are also the tasks that still need people, because no camera-only system handles them reliably.

The gap is at the point of contact

Watch a capable robot fail at a deformable task and you will see a pattern. The plan was fine. The motion was fine. Then something happened at the moment of contact.

The cable slipped mid-route. The connector half-seated. Two layers of film were picked as one.

None of those failures are visible from above. The latch is hidden behind the connector housing. Clear film glares or disappears. A slipping pouch looks fine right up until it falls. The gap is at the point of contact, and video can’t see it.

Force/torque sensing doesn’t close that gap either. A wrist force sensor tells you how hard the arm is pushing. It does not tell you where contact is happening on the fingers, or whether the object is starting to slide. That is a different signal, and it is the one that matters here.

What the human hand already knows

Your hand has about 17,000 mechanoreceptive units.² When a held object starts to slip, your grip corrects in about 74 milliseconds, faster than you could consciously react.³ You never decided to tighten your fingers. Your hand did it for you.

That loop, sense contact, adjust, keep going, is what robots are missing. It is also why people can route cables and pack bags all day, and robots mostly can’t.

Other teams are already seeing what touch does

We are not the only ones who think touch matters. Published work from other groups points the same way:

• Liu et al. reported that adding touch took peg insertion success from 45% to 90% in their experiments.⁴
• Funk et al. reported robotic match-lighting going from 20% to 80% with touch.⁵
• Figure moved its Helix policy from vision-only to fingertip touch.⁶

These are their results, on their tasks, with their systems. They are not Cerebel results. But together they make a clear case: when a task depends on contact, giving the robot a sense of contact helps.

Where Cerebel fits

Frontier models decide what to do. Cerebel teaches how to manipulate using touch.

We are building the dexterity layer between frontier models and robot hands. A large model can choose to route a cable to port 3. Our job is the part after that decision: feeling the cable’s tension, catching the slip, confirming the click.

Concretely, we are working on three things.

A robot to work on. AdiBot is our mobile manipulator, assembled in Dallas. Version 1.0 is built: a 7-DoF arm, a lift, a gripper, depth cameras and teleoperation. The next generation is in development, with a dual-arm torso and dexterous hands designed to carry 48 tactile channels per finger, sampled at 100 Hz.

A model that treats touch as a first-class input. Our VLSA model (vision, language, sensory, action) pairs an open vision-language backbone with a Cerebel touch encoder that turns tactile signals into tokens, plus an action expert. It is designed to run closed loop at 100 Hz on the robot.

A way to capture skill. Tacto is a tactile capture system that records how skilled hands handle deformables.

We are a research-stage company, founded in October 2025 in Dallas–Fort Worth. Some of this is built. Much of it is in development. We will be clear about which is which, and we will publish what we learn as we test it.

Why start with things that bend

Deformables are the hardest test of touch, which makes them the right place to start. If a robot can feel a cable under tension or a second layer of film, it can feel a peg going into a hole or a box starting to slide.

Deformables are the start. The same touch intelligence transfers to any robot hand.

For an operations leader, that means the tasks you still staff by hand, the ones nobody has automated, are the ones we are aiming at. For a robotics engineer, it means a new sensory channel in the policy, not just a better camera. For an investor, it means a layer that sits underneath many robots rather than a bet on one.

Robots have learned to see and to plan. The next step is learning to feel. That is the work, and it starts with everything that bends.