Approaching The Augmentation of Heuristic Behaviors With Reinforcement Learning In Collective Robotic Construction

Architectural design in collective robotic construction (CRC) has been explored using various approaches. One method for exploring what robot swarms can assemble is agent-based modeling and simulation (ABMS), where each robot acts as an agent with specific behaviors contributing to the assembly process. ABMS in CRC, however, has generally been utilized to pursue emergent outcomes, in which the negotiation and iteration of heuristic behaviors result in varying design outcomes. This paper investigates how to design behaviors if the architectural goal is known but the actions to get there are not. Specifically, reinforcement learning is utilized on a two-wheeled robotic system to develop and evaluate a workflow for integrating learned and heuristic behaviors in the same model. By giving the robots higher intelligence through the combination of computational methods, the aim is that the machines in CRC can truly be deployed anywhere and still be able to achieve their construction goals.