The rapidly growing e-grocery sector demands fulfillment centers (FCs) that operate with high efficiency and robustness, even under strong variability in task durations. Traditional scheduling approaches, such as the Earliest Due Date–First Served (EDD-FS) heuristic, lack real-time adaptability and neglect uncertainty. This study develops a real-time adaptive scheduling framework that combines predictive analytics with optimization to improve the timeliness and reliability of picking operations in a large-scale e-grocery FC. Gradient Boosted Decision Tree (GBDT) models are trained to predict active and inactive batch picking durations, significantly outperforming baseline methods. These predictions feed into two Mixed-Integer Linear Programming (MILP) schedulers: a deterministic variant using point estimates, and a scenario-based variant that incorporates uncertainty via sampled task durations. Evaluated in Picnic’s proprietary warehouse simulation system, the deterministic scheduler reduces total delay by 16% and delayed batches by 8.5%, though performance depends on the EDD-FS weighting. The scenario-based scheduler achieves further improvements, decreasing total delay by 6.3% to 25.6% and maximum delay by 16.6% to 37.9%, with higher scenario counts improving outcomes at the cost of computation. These results show that integrating predictive models with deterministic and scenario-based optimization enhances schedule robustness compared to heuristic baselines. Future work should validate the framework in live FC operations, explore rolling-horizon and metaheuristic extensions, and assess reinforcement learning as an alternative for adaptive decision-making. This study thus provides a novel methodological framework with both academic significance and direct operational relevance for the rapidly expanding e-grocery sector.