Sensory-Motor Learning for Simultaneous Control of Motion and Force: Generating Rubbing Motion against Uneven Object

Abstract

We propose a motion generation model for simultaneous control of motion and force using deep learning. Conventional force control methods require expensive torque sensors and complex control theory, and implementing force control for each task requires huge development costs. In this paper, we realize rubbing motions against an uneven object at low cost by using a motion generation method that takes as input the joint angles and current values of an inexpensive servo motor. We evaluated the generalization ability of the model by confirming that the robot can perform rubbing motions against unlearned uneven or tilted objects. In addition, by comparing several motion generation models, we clarified that the following two components are important for simultaneous control of motion and force. (1) The joint angles and the current values are input to different neuron layers to extract the features. A time constant, which is the speed of information transfer, is set for each layer in order to integrate and learn input information with different time characteristics. (2) In the output part of the model, a single neuron layer is used to predict the joint angle and current value simultaneously. This makes it easy to extract features and integrate learning from two inputs with different time characteristics, and the robot can generate appropriate motions based on the contact situation in real time.