Online optimization for the locomotion of Roombots structures



This semester project focuses on the locomotion of Roombots, a modular robot designed at the Biorobotic Laboratory. It is a continuation of previous works by Soha Pouya, Jesse van den Kieboom and Alexander Spröwitz, who were able to found good locomotion patterns of several structures using a simulation in Webots. Yet, these optimization patterns lead to poor result when programmed on the physical robot. The main objective of this project is to repeat this optimization process on the physical robot (online optimization), in order to compare the results with the simulation (offline optimization). Such study has already been carried by Alexander Spröwitz, Rico Möckel and Jerome Maye, on a simpler modular robot (YaMoR).

Only one simple structure was considered, which consists of a metamodule (two connected module) in PER configuration. The robot is controlled throught a Central Pattern Generator, a biologically-inspired model responsible for the generation of rythmic movements of the motors.

At first, two optimization methods are compared: Particle Swarm Optimization and Powell’s method. The former turns out to be more robust against local minimas than the latter, and was chosen to be implemented on the controller. Then, the controller was improved to contain the optimization framework. A collision detector from Stephane Bonardi has also been added to avoid harmful movements. Finally, several experiments have been done, and have lead to some interesting locomotion pattern, which are significatively different from what was obtained with the simulation.