Multirobot Collaborative Pursuit Target Robot by Improved MADDPG.

Area of Science:

• Robotics
• Artificial Intelligence
• Reinforcement Learning

Background:

• Multirobot systems face policy formulation challenges, particularly in pursuit-evasion scenarios with sparse rewards and dynamic environments.
• Current decision-making methods relying solely on environmental rewards show limitations in achieving optimal strategies.
• Sparse rewards and environmental unpredictability hinder effective strategy learning in multirobot pursuit tasks.

Purpose of the Study:

• To propose an enhanced multirobot pursuit method using an improved multiagent deep deterministic policy gradient (MADDPG).
• To address the challenge of sparse rewards in pursuit-evasion scenarios by integrating intrinsic and external environmental rewards.
• To improve the learning efficiency and strategy development of multirobot agents.

Main Methods:

• Implementation of an improved multiagent deep deterministic policy gradient (MADDPG) algorithm.
• Integration of intrinsic rewards, generated by an intrinsic curiosity module with a state similarity module, alongside external environmental rewards.
• Utilizing a state similarity module with threshold constraints to balance exploration and exploitation for intrinsic reward effectiveness.

Main Results:

• Significant improvement in robot reward values and pursuit task success rates.
• Demonstrated ability of pursuers to close in on escapees more rapidly.
• Reduction in the average following distance between pursuers and escapees.

Conclusions:

• The proposed MADDPG-based method effectively enhances multirobot pursuit strategies by overcoming sparse reward limitations.
• Combining intrinsic and environmental rewards leads to more efficient learning and improved task performance.
• The method shows practical improvements in pursuit dynamics, reducing distances and increasing success rates.