Experimental demonstration of mutually coupled short-cavity DFB laser photonic reservoir computing for two-joint robotic manipulator dynamics prediction

To address the latency and power consumption bottlenecks of traditional digital methods in high-speed industrial robotic arm dynamics prediction, as well as the limitations of existing photonic reservoir computing (PRC) that primarily focuses on abstract time-series benchmarks and lacks hardware verification in engineering scenarios, this paper constructs a PRC experimental platform based on mutually coupled short-cavity distributed feedback (DFB) lasers with self-feedback. A dimension-wise normalization, fixed-order serial flattening, and time-multiplexed mask encoding scheme is proposed to efficiently map the 8-dimensional coupled state sequence of a two-joint robotic arm into a single-channel optical injection signal. The prediction is accomplished by exclusively training a regularized linear readout layer. Experimental results demonstrate that the normalized mean squared errors ( NMSE exp ) for predicting the end-effector across 9 types of regular and chaotic trajectories are all below 0.056. Furthermore, the mutually coupled parallel architecture reduces the average prediction error by 4.7% compared to a single reservoir setup, maintaining robust performance under variations in time-multiplexing parameters, critical operational parameters, and up to a 100% deviation in the robotic link lengths. This work establishes a reproducible experimental coupling between robotic arm dynamics prediction and real photonic hardware, providing a solid experimental foundation for the optical implementation of state prediction in high-speed robotic arms.