Multi-Agent Self-Assembly using Reinforcement Learning

Abstract

We explore how large systems of multiple agents can learn to assemble structures using Reinforcement Learning. Coordination in this setting is challenging because of the sparse nature of the task signal and because agents have to simultaneously react to a large number of constantly evolving peers. We investigate how to use contrastive learning to de-sparsify the setting from a binary signal to probability densities over states, resulting in intriguing scalability and generalization properties. We additionally show how constrained sensing can help us understand how much and in what ways agents need to perceive in order to coordinate effectively