Memetic Robot Control Evolution and Adaption to Reality

Inspired by animals’ ability to learn and adapt to changes in their environment during life, hybrid evolutionary algorithms have been developed and successfully applied in a number of research areas. This paper explores the effects of learning combined with a genetic algorithm to evolve control system parameters for a four-legged robot. Here, learning corresponds to the application of a local search algorithm on individuals during evolution. Two types of learning were implemented and tested, i.e. Baldwinian and Lamarckian learning. On the direct results from evolution in simulation, Lamarckian learning showed promising results, with a significant increase in final fitness compared with the results from evolution without learning. Further experiments with learning on the real robot demonstrated an efficient adaptation of the robot gait to the real world environment, and increased the performance to the level measured in simulation. This paper demonstrates that Lamarckian evolution is effective in improving the performance of robot controller evolution, and that the same learning process on the physical robot efficiently reduces the negative impact of the simulation-reality gap. © IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works