Actuator-level motion and contact episode learning and classification using adaptive resonance theory

Abstract Several methods exist to detect and distinguish collisions of robotic systems with their environment, since this information is a critical dependency of many tasks. These methods are prevalently based on thresholds in combination with filters, models, or offline trained machine learning models. To improve the adaptation and thereby enable a more autonomous operation of robots in new environments, this work evaluates the applicability of an incremental learning approach. The method addresses online learning and recognition of motion and contact episodes of robotic systems from proprioceptive sensor data using machine learning. The objective is to learn new category templates representing previously encountered situations of the actuators and improve them based on newly gathered similar data. This is achieved using an artificial neural network based on adaptive resonance theory (ART). The input samples from the robot’s actuator measurements are preprocessed into frequency spectra. This enables the ART neural network to learn incrementally recurring episodic patterns from these preprocessed data. An evaluation based on preliminary experimental data from a grasping motion of a humanoid robot’s arm encountering contacts is presented and suggests that this is a promising approach.