Hybrid Rack–Machine-Learning Pointing Correction for a 1-m Robotic Telescope

We present a hybrid pointing-correction framework for robotic telescopes and evaluate it under strict temporal generalization using real Alt--Az guiding data from a 1-m telescope in the Las Cumbres Observatory network. The framework combines a 19-parameter rack model as a physically motivated analytic backbone with an XGBoost-based residual regressor that learns the remaining error from rack-term contributions, Alt--Az geometry, and a compact set of environmental features. The primary evaluation uses rolling temporal splits, so that each test segment is strictly later than the data used for training and model selection. Under this protocol, the hybrid strategy achieves the best overall future-window pointing accuracy among the tested model families. Its advantage is clear relative to the purely analytic and purely data-driven baselines, while the difference from a simpler residual-learning baseline is smaller. In terms of weighted mean VecRMS, the corresponding values are 12.75~arcsec for Hybrid, 14.95~arcsec for Rack+ResTree, 18.14~arcsec for Rack-only, and 21.87~arcsec for PureML. Additional drift diagnostics provide descriptive evidence of temporal distribution shift over the studied interval, supporting the use of time-aware evaluation. A strictly causal offline replay also suggests that a conservative validation-gated model-refresh policy may improve future-window pointing accuracy under controlled conditions. Additional analyses of condition-wise robustness, feature relevance, and computational overhead support the practical feasibility of the framework on the present dataset. These findings are limited to the studied telescope, feature set, and observation period; longer temporal baselines, cross-telescope validation, and controlled live deployment remain future work.