Robot Learning and Self-Sufficiency: What the energy-level can tell us about a robot’s performance

Abstract

Electrical energy is an important factor for a robotic agent when it tries to stay autonomously operational for some time. Monitoring this variable can provide important feedback for learning. In this paper, we present two different learning criteria based on this idea. Dealing with self-sufficient agents, i.e., agents that roughly speaking have a job to do, one criterion works over cycles of iterated “work” and “recovery”. In doing so, it gives some kind of feedback of the robot’s efficiency. We argue that a second criterion is needed for learning of most basic behaviors as well as in emergency situations. In these cases, fast and strong feedback, somehow comparable to pain, is necessary. For this purpose, changes in the short-term energy-consumption are monitored. Results are presented were basic behaviors of a robot in a a real-world set-up are learned using a combination of both criteria. Namely, the robot learns a set so-called couplings, i.e., combinations of simple sensor-processes with elementary effector-functions. The couplings learned enable the robot to do touch-based as well as active IR obstacle-avoidance and autonomous recharging on basis of phototaxis.