Demonstrating Targeted Navigation and Emergent Behaviour in Turbulent Flows via Deep Reinforcement Learning.

Understanding navigation in the presence of a turbulent flow has many applications including designing autonomous drones, creating targeted drug delivery technologies, and modelling weather patterns. Given the unpredictable nature of turbulence, it is challenging to build control schemes for efficient navigation. This study considers autonomous navigation in turbulent flows by training deep reinforcement learning algorithms in computer simulations of turbulent flows. Our swimmers learn to navigate a two-dimensional flow, and to exploit vortices in the fluid field to traverse their environments. We demonstrate that the RL based swimmers outperform simpler control theory-based approaches. We build on prior work by demonstrating emergent multi-agent behavior in independently trained swimmers, and showcase their ability to 'slipstream' each other in the flows to further optimize their movements.