Benefits of Social Learning in Physical Robots

Abstract

Robot-to-robot learning, a specific case of social learning in robotics, enables the ability to transfer robot controllers directly from one robot to another. Previous studies showed that the exchange of controller information can increase learning speed and performance. However, most of these studies have been performed in simulation, where robots are identical. Therefore, the results do not necessarily transfer to a real environment, where each robot is unique per definition due to the random differences in hardware. In this paper, we investigate the effect of exchanging controller information, on top of individual learning, in a group of Thymio II robots for two tasks: obstacle avoidance and foraging. The controllers of the robots are neural networks that evolve using a modified version of the state-of-the-art NEAT algorithm, called cNEAT, which allows the conversion of innovations numbers from other robots. This paper shows that robot-to-robot learning seems to at least parallelise the search, reducing wall clock time. Additionally, controllers are less complex, resulting in a smaller search space.