A Reinforcement Learning Approach to Multi-Robot Planar Construction

We consider the problem of shape formation in a decentralized swarm of robots trained with reinforcement learning. Shapes are formed from ambient objects which are pushed into a desired pattern. The shape is specified using a projected scalar field that the robots can locally sample. This scalar field plays a similar role to the pheromone gradients used by social insects such as ants and termites to guide the construction of their sophisticated nests. The overall approach is inspired by our previously developed orbital construction algorithm. In this paper, we use reinforcement learning to automatically learn policies that accomplish shape formation without the need for hand-coding algorithmic solutions for each desired shape. The particular research questions addressed in this paper are as follows: (1) The performance of learned policies versus the original hard-coded orbital construction algorithm; (2) The performance of the system on more shapes than were considered for the original algorithm. (3) The impact of the number of robots used in training and then subsequently in testing; We provide experimental results using a custom two-dimensional physics simulator of an environment containing circular robots and objects.