BlogTechnology4 min read
1D, 2D, 3D: a field guide to deformable objects
Cables, bags and foam all bend, but they bend in different ways. A short guide to the three families of deformable objects and why each one is hard for robots.
Most of the objects robots have learned to handle are rigid. A box is a box whether you grab it from the top or the side. Most of the objects people handle at work are not like that. They are cables, bags, film, fabric, foam and pouches. They bend, fold, stretch and squash. Engineers call them deformable objects, and they remain an open research problem. This is a short guide to how we think about them.
Why “deformable” is its own category
A rigid object has a fixed shape. If a robot knows where it is and how it is oriented, it knows nearly everything it needs to pick it up.
A deformable object breaks that assumption in three ways:
- It changes shape when you touch it. The act of grasping changes the object you are grasping.
- It hides its own state. A camera can show you the outline of a cable. It can’t show you the tension inside it.
- It is hard to model ahead of time. The same bag behaves differently depending on how full it is, how it was folded and how it was picked up.
These are the tasks Moravec’s paradox predicts would be hard: easy for a person, very hard for a machine.3
We sort deformables into three families, by shape.
1D: wires, cables and hoses
What they are. Long, thin objects whose length dwarfs their width. Power cables, data cables, wire harnesses, tubing, hoses.
Where you find them. Server racks, vehicle assembly, appliance manufacturing, electrical panels, anywhere liquid or power has to be routed.
What makes them hard. A 1D object can be in an almost infinite number of shapes. It coils, twists and tangles. It carries tension along its length, which a camera can’t see. It can snag behind a bracket out of view. And at the end of most cables is a connector that has to be seated, where the latch is hidden inside the housing.
What touch adds. Tension in the grip as the cable is pulled. A sudden change in resistance when it snags. The sharp force drop of a connector clicking home.
Typical failure if you miss it. A strained or pinched cable, or a half-seated connector that works at first and fails later.
2D: bags, film and fabric
What they are. Flat, thin objects that fold and drape. Poly bags, shrink film, labels, garments, cloth.
Where you find them. Fulfillment and packing, apparel, food packaging, medical supply kitting.
What makes them hard. A 2D object can fold over itself, hiding most of its surface. Layers stick together. Clear film either glares under light or disappears entirely. A camera often can’t tell whether there is one layer or two in the gripper.
What touch adds. Thickness and friction between layers, which tell you how many you are holding. The feel of a layer starting to slide.
Typical failure if you miss it. Two bags picked as one, or a torn bag.
3D: foam, pouches and soft goods
What they are. Bulky objects that compress, sag or shift. Foam inserts, filled pouches, padded envelopes, soft produce, plush items.
Where you find them. Kitting, retail fulfillment, food handling, protective packaging.
What makes them hard. A 3D soft object looks similar whether it is pressed lightly or crushed. Its contents shift as it moves. A pouch can begin to slip long before anything is visible.
What touch adds. Pressure spread across the palm and fingers, which separates seating from crushing. Small shear at the fingertip that warns of slip early enough to re-grip.
Typical failure if you miss it. Crushed product and a rejected kit, or dropped items and a stopped cell.
A quick reference
| Family | Examples | Hidden state | What touch reads |
|---|---|---|---|
| 1D | Wires, cables, hoses | Tension, snags, latch engagement | Pull, resistance changes, the click |
| 2D | Bags, film, fabric | Layer count, sliding | Thickness, friction, slip |
| 3D | Foam, pouches, soft goods | Compression, shifting contents | Pressure distribution, fingertip slip |
The common thread
Across all three families, the information that decides success sits at the point of contact. People handle it without thinking. The human hand has about 17,000 mechanoreceptive units,1 and corrects a slipping grip in about 74 milliseconds.2 That is fast enough to catch a slip before it becomes a drop.
Robots trained mostly on camera data don’t have that channel. A wrist force/torque sensor helps a little, but it measures how hard the arm presses in total, not where contact is or whether something is sliding. And open manipulation datasets carry almost no tactile signal, so even a well-trained model has rarely been shown what slip or a click feels like.
Why we organise our work this way
At Cerebel, we use these three families to structure what we build and test. Each one stresses touch in a different way: 1D on tension and seating, 2D on layers and friction, 3D on pressure and slip. A system that handles all three has to read touch well, not just add a sensor.
Our next-generation AdiBot hands are designed with 48 tactile channels per finger at 100 Hz, and our VLSA model is designed to turn those signals into touch tokens for the policy. We are testing across all three families. We will share results as we have them.
Deformables are the start. The same touch intelligence transfers to any robot hand. If a robot can tell one layer of film from two, it can tell a lot of other things it currently can’t.