Evolving Error Tolerance in Biologically-Inspired iAnt Robots

Abstract

Evolutionary algorithms can adapt the behavior of individuals to maximize the fitness of cooperative multi-agent teams. We use a genetic algorithm (GA) to optimize behavior in a team of simulated robots that mimic foraging ants, then transfer the evolved behaviors into physical iAnt robots. We introduce positional and resource detection error models into our simulation to characterize the empirically-measured sensor error in our physical robots. Physical and simulated robots that live in a world with error and use parameters adapted specifically for an error-prone world perform better than robots in the same error-prone world using parameters adapted for an error-free world. Additionally, teams of robots in error-adapted simulations collect resources at the same rate as the physical robots. Our approach extends state-of-theart biologically-inspired robotics, evolving high-level behaviors that are robust to sensor error and meaningful for phenotypic analysis. This work demonstrates the utility of employing evolutionary methods to optimize the performance of distributed robot teams in unknown environments.