Experience-based generation of maintenance and achievement goals on a mobile robot

Abstract

Learning skills or knowledge online from experiences is attractive for robots because it permits them to develop new behavior autonomously. However, the onus lies with the system designer to specify which skills or knowledge the robot should learn. Experience-based goal generation algorithms permit a robot to decide autonomously what it will to learn. This paper presents an adaptive resonance theory approach to experience-based generation of approach, avoidance, maintenance and achievement goals for a mobile robot. An experimental analysis is conducted to explore the relationship between algorithm parameters and goals generated on a simulated ePuck robot. Results show how parameter choice influences the number, stability and nature of generated goals. We identify the weight representations, distance functions and update rules that are appropriate for a mobile robot to generate maintenance and achievement goals.