Abstract
Ken Larpin, Master project
The objective of this project is the co-evolution of virtual robots to make them perform various manoeuvrability tasks.
A virtual framework is implemented to produce robots with different morphologies and controls. A Multi-Objective Evolutionary Algorithm (MOEA) is then used to co-evolve the robots body and mind. The robots design is made to produce realistic legged robots with a forced symmetry along the spinal cord. This morphological constrain successfully made the robots converge toward efficient legged locomotion.
One robot is composed of a torso and various pairs of symmetrical legs. The torso and the legs are composed of different basic parts linked together by rotational actuators. The actuated Degrees of Freedom (DOF) are controlled by a Central Pattern Generator (CPG) network.
During evolution the basic parts composing a robot can be removed or new ones can be added. The geometrical sizes of the parts can mutate within some fixed values. The CPG network’s values can also mutate. Finally crossover can be applied to mix two genomes.
This framework is then used to co-evolve robots to perform two manoeuvrability tasks, steering and speed control. During those experiments a command is applied to the robot and modify some predefined parameters in the CPG network. A fitness is used during evolution in order to make the robots fit the command applied to them. The results has shown that co-evolution of morphology and control increase the manoeuvrability performances of the robots.