A Multi-Brain Approach for Multi-Tasking in Evolvable Robots

EasyChair Preprint 9881

Abstract

Simultaneously evolving morphologies (bodies) and controllers (brains) of robots for a specific task (objective) is challenging because it is not intuitively clear how the perfect morphology looks given a certain task and environment. Even when a morphology is found and fixed, learning its affiliated controller to perform well is a second step of elaborate experimentation. Naturally, the complexity is increasing with the number of objectives that the robot is supposed to solve. In this paper we are proposing an approach to tackle this problem in a multi-objective setup by showing an efficient way to achieve both goals, finding a good phenotype for a given task and environment and getting a controller that can be used to complete the task with this phenotype. To this end, we show how to solve the bi-objective optimization problem of rotating to search for a given target and then moving to its position. We find this to be a rather useful skill for both robots and animals and it is an analogy to the real-world need of search and hunt for food or other targets. To achieve it, we apply the NSGA-II algorithm for a multi-objective selection and to create a population of appropriate morphologies with plastic brains. Out of this population the most promising candidates are selected and two randomly initialized brains are evolved separately for the given tasks, that can then be used by a state machine to switch between the behaviors.

Keyphrases: NSGA-II, Targeted Locomotion, evolutionary robotics, multi-objective optimization

@booklet{EasyChair:9881,
  author    = {Ege de Bruin and Julian Hatzky and Babak Hosseinkhani Kargar and A.E. Eiben},
  title     = {A Multi-Brain Approach for Multi-Tasking in Evolvable Robots},
  howpublished = {EasyChair Preprint 9881},
  year      = {EasyChair, 2023}}