Embedded Files
Goal: Our objective is to teach a robot arm to do contact rich manipulation: 1) pick up a peg and 2) place the peg in a hole. We used Diffusion EDF for our long-horizon policy because its SE3 equivariant properties allow it to be highly data-efficient, learning tasks from only 10-15 short demonstrations, as well as being generalizable to new configurations of target objects in the scene. Diffusion EDF was previously designed for pick and place tasks only, so we aimed to assess its performance for the peg in hole insertion task.
The Peg-in-Hole is the precursor to more complex industrial assembly tasks. It is a contact rich task that requires significantly more precision than traditional pick-place problems and handling the forces between target components requires more capable control algorithms. To this end, we augment our open loop vision model with a geometric impedance control algorithm to control force and position at the end effector simultaneously.
Developing pipelines to accomplish a peg-in-hole task could pave the path to robotic platforms capable of more challenging assembly tasks in a factory or construction setting. Such robotic platforms can be used to reduce construction time or perhaps even create structures in environments unreachable to humans (in extraterrestial settings). In addition, robots capable of assembly tasks in more unstructured settings also allows setting up factories of varying scales quickly and more affordably.
Page updated
Report abuse