Data-Model Hybrid-Driven Safe Reinforcement Learning for Adaptive Avoidance Control Against Unsafe Moving Zones.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Machine Learning

Background:

• Safety is a critical concern in the expanding application of reinforcement learning (RL).
• Existing methods struggle with avoidance control in dynamic environments with multiple moving unsafe zones.

Purpose of the Study:

• To develop a novel data-model hybrid-driven safe RL (SRL) scheme for effective avoidance control.
• To address the challenge of operating in domains with multiple, moving unsafe zones.

Main Methods:

• Encoding a barrier function (BF) into the cost function to transform the avoidance problem into an optimal control problem.
• Utilizing integral RL (IRL) with an actor-critic neural network (NN) structure and a state-following (StaF) kernel function for adaptive policy generation.
• Employing a state extrapolation technique to integrate both real-time and simulated experience for policy learning.

Main Results:

• Theoretical substantiation of closed-loop stability and weight convergence for the proposed SRL scheme.
• Demonstrated effectiveness on single integrator, nonlinear numerical, and unicycle kinematic systems.
• Highlighted advantages over existing control methods through comparative analysis.

Conclusions:

• The proposed data-model hybrid-driven SRL scheme provides a robust solution for avoidance control in complex, dynamic environments.
• The method ensures safety and stability while achieving effective control policy learning.
• This work advances the practical application of safe RL in real-world control scenarios.