Efficient collective swimming by harnessing vortices through deep reinforcement learning

Area of Science:

• Fluid dynamics
• Biomechanics
• Robotics

Background:

• Fish schooling is observed to provide evolutionary advantages, including energy conservation.
• The precise physical mechanisms behind energy savings in fish schools remain largely unknown.
• Understanding these mechanisms could inform bio-inspired robotic systems.

Purpose of the Study:

• To investigate the physical mechanisms by which fish achieve energy savings while schooling.
• To determine optimal strategies for fish to exploit the vortical wakes of conspecifics.
• To explore the application of deep reinforcement learning in simulating and understanding fluid-structure interactions.

Main Methods:

• High-fidelity computational fluid dynamics (CFD) simulations were employed.
• A deep reinforcement learning (RL) algorithm, utilizing recurrent neural networks with long-short-term memory (LSTM) cells, was developed.
• The RL agent learned navigation policies within complex, unsteady, vortical flow fields.

Main Results:

• Fish can enhance propulsive efficiency by positioning themselves optimally within the vortex wakes of other swimmers.
• Off-center positioning relative to leaders, not in-line swimming, yields energetic benefits for followers.
• Fish synchronize body deformation with oncoming vortices to harvest energy, improving efficiency without increasing the leader's energetic cost.

Conclusions:

• Fish actively harvest energy from vortices shed by conspecifics, confirming the energetic advantages of schooling.
• This study demonstrates the potential of deep RL for developing navigation strategies in complex fluid environments.
• Findings have implications for improving energy efficiency in autonomous robotic swarms operating in fluid dynamics.