Shattering Latency Boundaries: An LSTM–RCM Driven Smith Predictor for Path Tracking Control Under Time Delay

Time-varying communication delays fundamentally limit the performance of networked control systems in robotic applications. Classical Smith predictors can compensate for constant delays but degrade under nonlinear dynamics and stochastic variations. This article presents a predictive delay-compensation framework that embeds a long short-term memory (LSTM) dynamics model and a sparse variational Gaussian process (SVGP) residual corrector referred to as a residual correction model (RCM) within a Smith predictor architecture, bridging classical control with data-driven modeling. The LSTM forecasts delay-free state evolution, while the SVGP refines these predictions by modeling structured residuals with quantified uncertainty. The approach was validated on a mobile robot teleoperation platform executing straight, circular, and mixed trajectories under round-trip delays of up to 1 s. Across all tests, the LSTM–RCM Smith predictor achieved lower lateral and heading errors than a pure pursuit controller without delay compensation, the conventional Smith predictor, and an enhanced Smith predictor with$\boldsymbol{H}$-filter compensation, demonstrating its robustness and accuracy under stochastic network delays.