Multi-robot hierarchical safe reinforcement learning autonomous decision-making strategy based on uniformly ultimate...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Theory

Background:

• Deep reinforcement learning is crucial for multi-robot systems but struggles with safety in dynamic environments.
• Lack of security in autonomous decision-making exposes multi-robot systems to risks and damage.

Purpose of the Study:

• To develop a safety-focused reinforcement learning strategy for autonomous multi-robot systems.
• To address security challenges in dynamic and unstructured environments.

Main Methods:

• Proposed an event-triggered hierarchical safety reinforcement learning framework using constrained Markov decision processes.
• Integrated upper-tier evolutionary and lower-tier restoration networks for security and efficiency.
• Incorporated Lyapunov safety cost networks and Lagrange multipliers for strategy optimization.

Main Results:

• Demonstrated that action trajectories can be reverted to a safe space within finite time.
• The uniformly ultimately bounded constrained hierarchical safety reinforcement learning strategy (UBSRL) effectively restricts safety indicators below thresholds.
• Significantly enhanced stability and task completion rates in standardized scenarios.

Conclusions:

• The UBSRL strategy theoretically guarantees the efficacy of safety constraints for multi-robot systems.
• Achieved a balance between decision-making security and efficiency in autonomous systems.
• The approach provides a robust solution for safe navigation and task execution in complex environments.