Learning Multi-agent Workload Distributions in Confined Excavation

Our recent work with collective fire ant tunnel excavation revealed that approximately 30% of the workers performed 70% of the digging [Aguilar et al, Science 2018]. Complementary robot swarm experiments and numerical/theoretical models demonstrated the importance of this unequal workload distribution strategy in mitigating clogs and jams in congested environments. This contrasts with a strategy for spacious tunnels, in which maximal robot activity enables high performance. We have now augmented the robots with contact sensing capabilities so that each robot can distinguish different kinds of contact; preliminary results suggest that the duration and timing of robot-robot contacts correlates with tunnel density. We hypothesize that with the ability to deduce local tunnel density via contact sensing, the robots can excavate more effectively by adjusting their behavior transition probabilities based on their sensed environmental information. By varying the learning rules that rely on the local dynamics of the environment, we posit that the agents can learn transition probabilities to achieve an optimal digging performance. We explore the extent to which a learning-based approach can generate strategies that are robust to external and internal disturbances.