Detecting insertion tasks using convolutional neural networks during robot teaching-by-demonstration

Abstract

Today, collaborative robots are often taught new tasks through “teaching by demonstration” techniques rather than manual programming. This works well for many tasks; however, some tasks like precise tight-fitting insertions can be hard to recreate through exact position replays because they also involve forces and are highly affected by the robot's repeatability and the position of the object in the hand. As of yet there is no way to automatically detect when procedures to reduce position uncertainty should be used. In this paper, we present a new way to automatically detect insertion tasks during impedance control-based trajectory teaching. This is accomplished by recording the forces and torques applied by the operator and inputting these signals to a convolutional neural network. The convolutional neural network is used to extract important features of the spatio-temporal forces and torque signals for distinguishing insertion tasks. Eventually, this method could help robots understand the tasks they are taught at a higher level. They will not only be capable of a position-time replay of the task, but will also recognize the best strategy to apply in order to accomplish the task (in this case insertion). Our method was tested on data obtained from 886 experiments that were conducted on eight different in-hand objects. Results show that we can distinguish insertion tasks from pick-and-place tasks with an average accuracy of 82%.