Geometric reasoning for fine motion planning

In an assembly plan, a sequence of subtasks has to be determined, which require a lower level plan involving fine motion. The need for combining task-level knowledge and sensor-based information is unavoidable, since the use of force and torque sensor signals allows us to identify the contact state in the real world and verify whether the predictions of the task planner are correct or not. This paper builds upon previous results regarding error detection for plan monitoring. We extend them by deriving a geometric-reasoning world model for the peg-in-hole insertion task, and integrating it with a perception-based model obtained using neural networks. A novel learning scheme to identify contact states is also presented. As a result, an integrated approach to fine motion planning for assembly is developed, including perception, robotics and artificial intelligence techniques