Event-Triggered Safety-Stability Framework for Learning-Based Control of Multiagent System With Uncertain Dynamics.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Machine Learning

Background:

• Multiagent reinforcement learning (MARL) faces challenges in exploring high-dimensional spaces safely, risking system instability and unsafe states.
• Deployment in safety-critical systems is hindered by the need for robust safety guarantees and stability under uncertain dynamics.

Purpose of the Study:

• To propose a novel safety-stability layer for MARL systems operating in unknown environments with uncertain dynamics.
• To integrate robust control barrier functions (RCBFs) and input-to-state stable control Lyapunov functions (ISS-CLFs) for enhanced safety and stability.

Main Methods:

• A safety-stability layer is developed, integrating RCBFs and ISS-CLFs for multiagent systems.
• Safety-stability constraints are incorporated into the MARL framework, separating training (goal-reaching) and deployment (filtered policy outputs).
• An event-triggered mechanism for action compensation is designed to optimize computational resource usage based on safety assessments.

Main Results:

• The proposed method ensures strict adherence to safety constraints during policy execution.
• Task execution efficiency of multiagent systems is significantly enhanced.
• Validation through simulations in dynamic multiunicycle environments confirms the approach's effectiveness.

Conclusions:

• The integrated safety-stability layer effectively addresses exploration challenges in MARL for safety-critical applications.
• The method provides a robust solution for balancing exploration, safety, and efficiency in multiagent systems.
• The approach demonstrates practical applicability and potential for real-world deployment in complex robotic systems.