Deep Reinforcement Learning based Swarm Motion for Collision Avoidance via Self-configurable Potential Formation

One of the main challenges in swarm robotics is modeling and controlling the collective dynamics arising from local interactions among many agents. Classical control methods struggle with the nonlinear, decentralized, and adaptive behaviors required for swarm coordination in dynamic environments. This work presents a Deep Reinforcement Learning (DRL)-based flocking control framework that enables emergent dynamics within an adjustable formation while ensuring collision avoidance. A Deep Deterministic Policy Gradient (DDPG) agent learns coordinated strategies directly from experience, avoiding explicit modeling of inter-agent dynamics. To represent a formation-adaptable model, a Potential Linked Nodes (PLN) structure modifies swarm kinematics to reach goals and avoid obstacles. The DRL-based model regulates swarm positioning via Artificial Potential Fields (APF), while the agent controls collective motion through linear and angular velocities. Experiments with different swarm sizes and environments assessed performance using flocking m etrics such as distance error and collisions, confirming s table transitions. Results validated that the DDPG agent achieves reliable flocking toward targets while avoiding obstacles.