Photonic imitation learning for control tasks using deep photonic agentic networks

Abstract

We demonstrate a silicon photonic agentic platform for imitation learning using a 10-layer Mach–Zehnder interferometer network with random taper-induced phase shifts and nonlinear optical-to-electrical transformation. This architecture performs random linear projections with high throughput and low loss. Evaluated on two sequential control tasks (MountainCar momentum-based driving with 2-dimensional state inputs and Acrobot dual-pendulum with 6-dimensional state inputs), the system receives these states as optical amplitudes/phases and outputs 3 control actions via a trained linear readout. Our photonic network achieves 96.2% accuracy on MountainCar and 94.8% on Acrobot, significantly outperforming digital baselines (88.0% and 86.2%, respectively), and demonstrating its effectiveness for edge-compatible, low-power robotic control