Published
An industrial robot placing an engine block is working with an object that has one shape. That shape is known before the robot moves, it does not change while the robot holds it, and it will be the same shape when the robot lets go. Almost everything convenient about rigid-body manipulation follows from that single property.
Fabric has none of it. A folded piece of cloth and the same piece laid flat are the same object in different configurations, and the number of configurations is not finite in any useful sense. This is the gap between a problem that is largely solved in structured settings and one that remains genuinely open.
The state you cannot write down
Rigid-body pose is six numbers. A deformable object's configuration is, in principle, the position of every point on it — a continuous field rather than a pose. Any practical system approximates it, and every approximation discards something the next step might have needed.
The approximations that work tend to be task-specific. Grasping a towel by a corner needs corner detection, not a full shape model. Feeding fabric into a seam needs edge tracking. A general representation of cloth state that serves every downstream task is not currently available, and the useful systems are the ones that avoided needing one.
Dynamics that punish open-loop planning
A rigid object moves as the gripper moves it. Cloth does not. It drapes, it buckles, it sticks to itself, and the result depends on material properties that vary between batches of what the specification calls the same fabric.
This means a plan computed before the motion starts is a weaker guide than it is for rigid parts. Perception has to stay in the loop throughout the motion rather than only at the start, which raises the required sensing rate and makes the control problem harder in the same step.
Why simulation helps less than expected
Learning-based approaches to rigid manipulation lean heavily on simulation, because a simulator can produce vastly more interaction data than a physical cell. Cloth simulation is comparatively expensive and comparatively less faithful, so the gap between simulated and real behaviour is wider precisely where the data is most needed.
Work in this area tends to be honest about it: transferring a policy trained on simulated cloth to real cloth is an active research question rather than a matter of engineering effort.
What this implies for the near term
The practical position is that deformable handling is tractable where the geometry is bounded — a known material, presented in a known way, for a known operation — and open where it is not. That is a narrow window, but it is a real one, and it covers more of textile and assembly work than the general framing suggests.
It also argues for sequencing. Inspection of deformable materials is a comparatively well-posed vision problem and delivers value on its own. Handling is the harder half, and there is no particular reason to attempt both at once.