Online sensorimotor learning and adaptation for inverse dynamics control

Abstract

We propose a micro-data (< 10 trials) sensorimotor learning and adaptation (SEED) model for human-like arm inverse dynamics control. The SEED model consists of a feedforward Gaussian motor primitive (GATE) neural network and an adaptive feedback impedance (AIM) mechanism. Sensorimotor weights over trials are learned in the GATE network, while the AIM mechanism is used to online tune impedance gains in a trial. The model was validated by periodic and non-periodic tracking tasks on a two-joint robot arm. As a result, the proposed model enables the arm to stably learn the tasks within 10 trials, compared to thousands of trials required by state-of-art deep learning. This model facilitates the exploration of unknown arm dynamics, in which the elbow joint requires much less active control compared to the shoulder. This control goes below 3% of the overall effort. This finding complies with a proximal–distal control gradient in human arm control. Taken together, the proposed SEED model paves a way for implementing data-efficient sensorimotor learning and adaptation of human-like arm movement.