Towards realising multimodal robots

Evolutionary Algorithms (EAs) have been applied in co-evolutionary robotics over the last quarter century. They simultaneously generate the robot body morphology and controller algorithm through artificial evolution. However, the literature shows that in this area, it is still not possible to generate robots that perform better than conventional manual designs, even for simple tasks. This thesis describes steps undertaken to improve the co-evolution process. The investigation concerns two key areas of co-evolutionary robotics. The first is in fitness function development, which plays an integral part in selecting parents for mating during evolution. The effect of an incremental fitness function based on established algorithmic techniques from specific task domains of robotics is studied. An A-star algorithm-based fitness function for path planning is designed and implemented for co- evolution of robots for navigation and obstacle avoidance. Results show that the trajectories of robots that reach a goal using the A-star algorithm-based fitness function are shorter than the robots evolved with a basic distance-based fitness function. The second area of study is the role of the controller evolution in the co-evolution process. Inspired by natural paradigms of evolution coupled with learning in biological organisms, a Reinforced Co-evolutionary Algorithm (ReCoAl) is proposed. ReCoAl works by allowing a direct policy gradient based Reinforcement Learning algorithm to improve the controller of every evolved robot to better utilise the available morphological resources before the fitness evaluation. The findings indicate that the learning process has both positive and negative effects on the progress of evolution, similar to observations in evolutionary biology.

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

University of Greenwich

Year dc:date.issued

2019

Author dc:creator

• Radhakrishna Prabhu, Shanker Ganesh

Advisors dc:contributor.advisor

• Kyberd, J
• Seals, Richard

• Q Science (General)

Language dc:language

en