Real-World Validation of a Learned Policy for Cooperative Object Transport with the Transformer Encoder

In recent years, swarm robotic systems are expected to be used in open environments such as disaster sites. Considering applications in open environments, systems are required to maintain functionality even when the number of robots changes or communication among them is unstable. On the other hand, the design methodology for systems that take these constraints into account has not been sufficiently discussed. Our previous research proposed the use of the transformer encoder (self-attention mechanism), enabling the design of a neural network that remains effective even when the amount of observation data available to each robot varies. However, that work did not include validation in real-world environments, and it remains unclear whether the cooperative behaviors learned using the transformer encoder can be effectively applied outside of simulation. To address this, this paper aims to investigate the feasibility of sim-to-real transfer and examines the differences between the simulation and real-world environments. After training a policy model for cooperative object transport in a simulation environment, we conduct real-world experiments using TurtleBot3 Burgers to evaluate its applicability, adaptability to changes in robot numbers without retraining, and differences between simulation and real-world environments.