Learning Robot Control Using a Hierarchical SOM-Based Encoding

Abstract

Hierarchical representations and modeling of sensorimotor observations is a fundamental approach for the development of scalable robot control strategies. Previously, we introduced the novel Hierarchical Self-Organizing Map-based Encoding algorithm (HSOME) that is based on a computational model of infant cognition. Each layer is a temporally augmented SOM and every node updates a decaying activation value. The bottom level encodes sensori-motor instances while their temporal associations are hierarchically built on the layers above. In the past, HSOME has shown to support hierarchical encoding of sequential sensor-actuator observations both in abstract domains and real humanoid robots. Two novel features are presented here starting with the novel skill acquisition in the complex domain of learning a double tap tactile gesture between two humanoid robots. During reproduction, the robot can either perform a double tap or prioritize to receive a higher reward by performing a single tap instead. Secondly, HSOME has been extended to recall past observations and reproduce rhythmic patterns in the absence of input relevant to the joints by priming initially the reproduction of specific skills with an input. We also demonstrate in simulation how a complex behavior emerges from the automatic reuse of distinct oscillatory swimming demonstrations of a robotic salamander.

A Lyapunov-Stable Adaptive Method to Approximate Sensorimotor Models for Sensor-Based Control

“…These approaches require very large amounts of training data to properly constrain the learning algorithm, which is impractical in many situations. Other distributed implementations (based on SOM-like sensorimotor “patches,” Kohonen, 2013 ) are reported e.g., in Zahra and Navarro-Alarcon ( 2019 ), Pierris and Dahl ( 2017 ), and Escobar-Juarez et al ( 2016 ), yet, the stability properties of its algorithms are not rigorously analyzed.…”