When undertaking the basic task of mimicking ants as they collect food from remote sites to bring back to the nest, robot systems exhibit two types of errors: erroneous estimation of robot position and erroneous detection of a food resource. This paper proposes an evolutionary approach to managing these errors in a multiple robot system.
The authors have developed a central-place foraging algorithm (CPFA) based on ant foraging behavior. The behavior of each robot is defined by a number of parameters used in the CPFA, such as the probability of stopping the search effort, and a genetic algorithm (GA) is used to optimize behavior.
Two types of simulated worlds were used for the experiments: an error-free world (perfect world) and a world with error (imperfect world). Learned parameters from the perfect world were swapped with those in the imperfect world. The parameter swapping was also applied between two imperfect worlds with different grades of error. The results show that the GA can help the robot deal with sensor error, especially when the parameters are learned from an imperfect world and applied to a similar type of imperfect world.
Overall, it is interesting to see how biological evidence can help solve problems in a robotic system. The CPFA looks to be a promising approach to solving this foraging task. The main drawback of the paper is that all of the experiments were conducted in simulated worlds. Although two types of errors were taken into account, it is still far from reality. For example, the authors do not consider the detection error generated when the robot recognizes a laid pheromone (the landmark trace to a food location). Training in an imperfect world and testing in another imperfect world is not really novel. Another drawback is that there are no comparisons to other algorithms, although this may be due to constraints on the length of the report. It would be more convincing if the authors could add real robot experiments and comparisons.