Multi Robot Learning by Demonstration

Abstract

In this paper, we investigate the feasibility of a Multi Robot Learning by Demonstration system, which allows multiple teachers to give a demonstration to multiple robots simultaneously. A novel, complete end-to-end system was developed, which extracts data from a live human group demonstration, and allows the robots to imitate the demonstration by adapting the demonstration dataset to the current, possibly different environment. The complete system was evaluated using a series of increasingly difficult benchmark experiments, including a collaborative door opening experiment using a group of heterogeneous robots. The results showed, that the system is resistant to changes in the environment, as it was possible to give a demonstration in one environment, move the robots to a physically different but similar location, where the robots could still imitate the demonstration in this new context. The door opening experiment also shows that this system can be used to demonstrate and learn collaborative behaviour. Our results demonstrate a novel and promising method for teaching a group of robots to perform a joint task by human team demonstration.