Teaching robots pottery making tasks is easier said than done. This can be challenging for humans, doubly so for robots. Humans are capable of manipulating deformable objects in complex shapes because we can reason about material properties and figure things out in the presence of geometric occlusion. The RoPotter System explores a robotic pottery system that learns through demonstration.
The researchers used a gaming controller for end-effector positioning and orientation control. This system uses 2 x RGB-D cameras to observe clay deformation. 40 demonstrations were collected for a wide and tall bowl. A diffusion policy for each goal bowl shape was trained.
[HT]