An evolutionary approach for robust adaptation of robot behavior to sensor error

Abstract

Evolutionary algorithms can adapt the behavior of individual agents to maximize the fitness of populations of agents. We use a genetic algorithm (GA) to optimize behavior in a team of simulated robots that mimic foraging ants. We introduce positional and resource detection error models into this simulation, emulating the sensor error characterized by our physical iAnt robot platform. Increased positional error and detection error both decrease resource collection rates. However, they have different effects on GA behavior. Positional error causes the GA to reduce time spent searching for local resources and to reduce the likelihood of returning to locations where resources were previously found. Detection error causes the GA to select for more thorough local searching and a higher likelihood of communicating the location of found resources to other agents via pheromones. Agents that live in a world with error and use parameters evolved specifically for those worlds perform significantly better than agents in the same error-prone world using parameters evolved for an error-free world. This work demonstrates the utility of employing evolutionary methods to adapt robot behaviors that are robust to sensor errors.