Clustered Optimization of a Small-Scale Robot Swarm with Minimal On-Board Sensing

This paper explores the ability of a swarm of small-scale robots to optimize robot locomotion speed if only extremely limited on-board sensing capability is granted for each individual robot. This is done to assess benefits of swarm optimization versus individual control. The optimization of simulated robot locomotion speed is evaluated under multiple classical and machine learning algorithms. Furthermore, a new algorithm, Clustered Particle Swarm Optimization (CPSO), is developed based on existing optimization and clustering algorithms. The algorithms are based on an experimentally-validated nominal dynamic model for small-scale robots, with environmental parameters extended over a wider range of uncertainty. The nonlinear relationship between multiple actuation parameters and robot performance, specifically robot speed in this work, encourages a numerical method to search for a set of actuation parameters that leads to a higher robot speed. Three algorithms are presented and compared, including Simulated Annealing (SA), Particle Swarm Optimization (PSO), and CPSO. The comparison between SA and PSO demonstrates the utility of optimizing robots in a swarm. By increasing the number of robots, the optimizing iteration number can be reduced. The comparison of simulation of PSO and CPSO shows a 23% improvement with minimal sensing assumed from the robots. The single sensing feature, selected manually using phase plots, is the average foot height at a certain actuation phase in each cycle.