Robots Need Some Education: On the complexity of learning in evolutionary robotics

Abstract

Evolutionary Robotics and Robot Learning are two fields in robotics that aim to automatically optimize robot designs. The key difference between them lies in what is being optimized and the time scale involved. Evolutionary Robotics is a field that applies evolutionary computation techniques to evolve the morphologies or controllers, or both. Robot Learning, on the other hand, involves any learning technique aimed at optimizing a robot's controller in a given morphology. In terms of time scales, evolution occurs across multiple generations, whereas learning takes place within the `lifespan' of an individual robot. Integrating Robot Learning with Evolutionary Robotics requires the careful design of suitable learning algorithms in the context of evolutionary robotics. The effects of introducing learning into the evolutionary process are not well-understood and can thus be tricky. This thesis investigates these intricacies and presents several learning algorithms developed for an Evolutionary Robotics context.

Figures (47)

• Figure 1: Triangle of Life
• Figure 2: Real instances of the modules we simulate. From left to right: Core module with micro-controller and battery, active hinge/joint, and brick.
• Figure 3: a) A single CPG. b) The CPG network for our Spider, containing 8 CPGs (numbers) with 10 connections (blue lines).
• Figure 4: Flow chart describing our method for assessing learner performance on evolvable robots. Top left: the process starts with an M-ER experiment evolved for high speed (i.e. task performance) and novelty (shown as green dots indicating fitness in morphology space). From this experiment we select individuals from the last generation. The collection of robot pictures shows the initial test suite which is hand-picked. Both the last generation and the initial test suite can be represented in morphological space. Right: the test suite is extended by iteratively adding individuals from the last generation based on maximizing novelty. Bottom: the resulting 20 robots are used to compare different learning algorithms.
• Figure 5: Test suite of all 20 robots, with number of weights to be optimized in brackets $(N_w)$. The base set consisted of three hand-designed robots (Spider, Gecko and Snake), and two offspring of the Gecko and the Spider (BabyA and BabyB). The remaining set is selected from a population of evolved robots, using novelty search.