BlogPerspectives4 min read
Frontier models can reason about a cable. They still can't route one.
Today's models can see a cable, name it and plan its path. The failure comes later, at the moment the hand touches it. That is Moravec's paradox, still at work.
Show a frontier model a photo of a server tray and ask it how to route the cable on the left to the port on the right. It will give you a good answer. It will name the cable type, describe a path that avoids the fan housing, and warn you not to bend it too tightly. Hand that same plan to a robot and watch what happens. Often, the cable slips out of the gripper halfway along the route.
Frontier models can reason about a cable. They still can’t route one.
A paradox that hasn’t gone away
In 1988, Hans Moravec wrote down something roboticists had already noticed. The things that are hard for humans, like formal reasoning, turn out to be relatively easy for machines. The things that are easy for humans, like perception and movement, turn out to be very hard.1
Nearly four decades later, the paradox has shifted, but it hasn’t gone. Machines now see well. They plan well. They can even generate fluent motion from a text instruction. What they still can’t do is the part a child can: hold something soft and feel what it’s doing.
What models already do well
It is worth being fair to how far things have come. Given a cable task, a capable model can:
- See and name the cable. It can find it in a cluttered image and tell you what it is.
- Plan where it goes. It can describe the route, the order of steps and what to avoid.
- Generate the motion. It can produce arm trajectories that follow that plan.
Each of those used to be a research problem on its own. Today they are, increasingly, solved well enough to build on.
What breaks at contact
The failures come at one specific moment: when the hand meets the object.
The cable slips mid-route. The gripper closed on it. The plan was right. But a cable under tension pulls, and the grip wasn’t adjusted as the pull increased. By the time a camera sees the cable has moved, it is already out of the hand.
The connector half-seats. The robot pushed the connector into the port. From the camera’s view, it looks in place. But the latch never engaged. The connection will work for a while and then fail.
Two layers of film are picked as one. Clear plastic glares or vanishes under lighting. The robot grabs what looks like one sheet and lifts two.
In each case, the model’s reasoning was sound. The error happened in a place the model had no way to sense. The gap is at the point of contact, and video can’t see it.
Why more video won’t fix it
A natural response is to train on more video, at higher resolution, from more angles. That helps at the margins. It doesn’t address the core problem.
The information that matters is not in the image. Whether a latch has clicked is a sharp drop in force under the fingertip. Whether a cable is about to slip is a small shear signal across the finger pad. Whether you are holding one sheet of film or two is a matter of thickness and friction you feel, not see.
People don’t solve these by looking harder. When a held object starts to slip, your grip corrects in about 74 milliseconds, without conscious thought.2 That loop runs on touch.
Force/torque sensors at the wrist don’t fill the gap either. They measure how hard the arm is pressing. They don’t say where contact is on the fingers, or whether something is sliding. And the open datasets most robot models learn from are almost entirely camera and joint data, with almost no tactile signal at all.
The evidence from other labs
Published work from other teams supports the idea that touch fixes contact failures. Liu et al. reported peg insertion success going from 45% to 90% when touch was added.3 Funk et al. reported robotic match-lighting going from 20% to 80% with touch.4
Those are their results, not ours. Neither is a cable task. But both are contact-heavy, and both point the same direction.
Two layers, two jobs
We think of the problem as two layers with two jobs.
Frontier models decide what to do. Cerebel teaches how to manipulate using touch.
The model on top chooses the goal: route this cable to that port. The layer underneath handles the execution at contact: keep the grip as tension builds, feel for the click, notice the snag and back off.
That lower layer is what we are building. Our VLSA model (vision, language, sensory, action) is designed to sit on an open vision-language backbone and add a touch encoder that turns tactile signals into tokens the policy can use. An action expert turns all of that into motion. The whole loop is designed to run at 100 Hz on the robot, so it can react within the time a slip unfolds.
We are testing this now. We don’t yet have results to share on cable routing, and we won’t claim any until we do.
What this means in practice
If you run operations, you probably have tasks that look automatable on paper and keep failing in practice. Look closely and many of them fail at the same place: the moment of contact with something soft, flexible or hidden.
If you build robots, you have likely seen a policy that does everything right until the gripper closes. That last centimetre is where the remaining work is.
The reasoning problem is largely handled. The handling problem is not. That is where we have chosen to work.