Optimizing metachronal paddling with reinforcement learning at low Reynolds number

Area of Science:

• Fluid dynamics and biomechanics
• Robotics and artificial intelligence

Background:

• Metachronal paddling is a widespread swimming strategy involving coordinated limb oscillations.
• This rhythm is observed across diverse organisms and Reynolds numbers, suggesting evolutionary optimization.
• The efficiency and selection of metachronal rhythms at low Reynolds numbers remain underexplored.

Purpose of the Study:

• To investigate emergent limb coordination patterns using reinforcement learning in a simulated low Reynolds number swimmer.
• To determine if the metachronal rhythm is selected by an AI agent under varying physical constraints.
• To identify the most efficient and fastest swimming strokes generated by the AI.

Main Methods:

• A reinforcement learning agent was developed to control an elongated swimmer with multiple paddle pairs.
• Simulations were conducted at zero Reynolds number with fixed paddle spacings.
• The agent learned distinct coordination patterns based on the paddle spacing parameter.

Main Results:

• At tight paddle spacings, the agent learned a back-to-front metachronal wave-like stroke, mimicking biological swimmers.
• Wider paddle spacings resulted in qualitatively different, non-metachronal limb coordination patterns.
• Stroke efficiency was independent of paddle number, with the metachronal wave being consistently most efficient.

Conclusions:

• Reinforcement learning can replicate biological metachronal paddling under specific conditions.
• Paddle spacing is a critical factor influencing the selection of swimming coordination strategies.
• The back-to-front metachronal wave represents a highly efficient propulsion mechanism at low Reynolds numbers, irrespective of the number of paddles.