Verifiably stable nonlinear control with reinforcement-learned diffractive optical networks.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Control Systems Engineering

Background:

• Diffractive optical networks (DONs) excel at AI tasks like object recognition.
• Their potential for stable, continuous nonlinear control is largely unexplored.
• Conventional control strategies struggle with complex nonlinear dynamical systems.

Purpose of the Study:

• Introduce a novel framework for stability control of continuous nonlinear dynamical systems using DONs.
• Integrate reinforcement learning with Lyapunov conditions for guaranteed closed-loop stability.
• Address limitations of existing methods, such as behavior cloning's cumulative drift.

Main Methods:

• Developed a Lyapunov-constrained reinforcement learning diffractive-optical network (LC-RLDON) framework.
• Integrated reinforcement learning with differentiable Lyapunov conditions for policy optimization.
• Utilized a passive DON and a lightweight electronic linear layer for real-time optical Actor inference.

Main Results:

• LC-RLDON demonstrated superior performance in controlling underactuated rotary inverted pendulums.
• Achieved stable equilibrium in 2.8 seconds and recovery from disturbance in 2.1 seconds.
• Outperformed behavior cloning, which consistently failed to achieve stable control.

Conclusions:

• DONs can deliver real-time, formally safe control for continuous nonlinear systems.
• The LC-RLDON framework overcomes limitations of previous DON-based controllers.
• Paves the way for practical implementation in low-power, high-performance intelligent systems for robotics and autonomous vehicles.