Smart navigation through a rotating barrier: Deep reinforcement learning with application to size-based separation of...

Area of Science:

• Physics
• Robotics
• Artificial Intelligence

Background:

• Active Brownian particles (microagents) are model systems for self-propelled entities in fluids.
• Navigation in complex environments is a key challenge for microagents.
• Deep reinforcement learning (DRL) offers powerful tools for optimizing agent behavior.

Purpose of the Study:

• To investigate shortest-time navigation strategies for microagents through a rotating potential barrier.
• To determine if a rotating barrier can facilitate size-based sorting of microagents.
• To explore the effect of environmental noise on sorting efficiency.

Main Methods:

• Employing deep reinforcement learning, specifically the advantage actor-critic approach, to train microagents.
• Modeling the environment as a viscous fluid with a rotating Gaussian potential barrier.
• Quantifying particle sorting efficiency using specific separation measures.

Main Results:

• Demonstrated that a rotating potential barrier enables size-based sorting of microagents, unlike a static barrier.
• Showcased that microagents of different radii arrive at distinct average times, facilitating separation.
• Found that training microagents in a noisy background enhances the precision of size-based sorting.

Conclusions:

• DRL provides an effective framework for developing navigation strategies for microagents.
• Rotating potential barriers are a viable mechanism for achieving size-based particle separation.
• Environmental noise can be leveraged to improve the performance of microagent sorting systems.