A Policy-Guided Reinforcement Learning Method for Encirclement Control in Multiobstacle Environment.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Multiagent encirclement with obstacle avoidance (EMOCA) presents a significant challenge due to the difficulty in balancing target encirclement and collision avoidance.
• Existing methods struggle to efficiently train agents for complex, simultaneous tasks.

Purpose of the Study:

• To propose a novel reinforcement learning (RL) method, multiregulator-assisted RL for encirclement control (MRA-RLEC), to address the EMOCA problem.
• To enhance training efficiency through jump-start learning and curriculum learning (CL).

Main Methods:

• MRA-RLEC employs curriculum learning by dividing the encirclement task into progressively difficult subtasks.
• Multiple regulators are used to manage encirclement, obstacle avoidance, and policy execution transitions.
• A global encirclement reward decomposition (GERD) method is introduced to mitigate reward sparsity.
• A bidirectional communication protocol is designed to reduce communication overhead.

Main Results:

• Extensive experiments demonstrate the robustness and superiority of the proposed MRA-RLEC method.
• The method's practical applicability is validated in a ROS2/Gazebo simulation environment using an omnidirectional vehicle model.

Conclusions:

• MRA-RLEC effectively addresses the challenges of multiagent encirclement with multiobstacle collision avoidance.
• The proposed method offers enhanced training efficiency and robust performance for complex robotic control tasks.