Towards Safe Action Policies in Multi-robot Systems with Causal Reinforcement Learning

Causal reasoning is increasingly used in reinforcement learning to improve the learning process in several dimensions: efficacy of learned policies, efficiency of convergence, generalisation capabilities, safety and interpretability of behaviour. Robotic systems are no exception, but mostly cover single-robot tasks, whereas applications and methods of causal reasoning to reinforcement learning in multi-robot systems are still under-explored. In this paper, we take a first step in the direction of investigating the opportunities and challenges of applying explicit causal reinforcement learning in multi-robot systems. In particular, we measure the impact of a simple form of causal augmentation in three different simulated robotic scenarios requiring increasing degrees of cooperation among robots, and with different reinforcement learning algorithms exploiting various degrees of collaboration between them.