Human-Centric Order Picking: Performance Prediction and Robot Assignment at a Robotic Fulfillment Center

Problem definition: E-commerce giants scale up their order-picking operations by adopting robotic fulfillment centers (RFCs). In RFCs, automated guided vehicles transport movable shelf racks to pickers’ workstations, instead of having human pickers travel to pick items. Unfortunately, this apparent relief for pickers turns out to be a curse: Pickers become the bottleneck in the order-picking process. They undertake high-intensity, stationary, and repetitive tasks, which often cause both physical and mental health problems. To ease this tension, we collaborate with a major e-commerce firm to study how RFCs can improve picking efficiency by accounting for heterogeneous picker performance. Methodology/results: We propose a novel distributionally robust human-centric picking performance prediction (DHPP) model to forecast two critical metrics of picker performance: picking time and performance inconsistency. The DHPP model addresses distributional uncertainty by incorporating probabilistic constraints without requiring knowledge of the true underlying distribution. It leverages the empirical mean and covariance of random features that characterize picker behavior to hedge against worst-case prediction errors. We reformulate the DHPP model into a tractable second-order cone program. Using the predicted metrics, we then design a mixed 0–1 program to optimize the picker-order assignments. Managerial implications: Our computational study demonstrates that the DHPP model significantly outperforms state-of-the-art forecasting models in prediction accuracy. Our simulation, calibrated with real data from JD.com, shows that our strategy reduces the number of unfulfilled items by 14.2% and improves average pickers’ picking productivity by 7.5%. These improvements suggest significant welfare gains for pickers, increasing their income while helping alleviate stress and health issues. Funding: This work was supported by the National Natural Science Foundation of China [J. Luo was supported by Grant 72261008, Z. Wu was supported by Grants 72172027 and 72293563, Z. Hao was supported by Grants 72001036 and 72232001, and W. Qi was supported by Grants 72242106, 72521001, and 72188101]. Supplemental Material: The online appendices are available at https://doi.org/10.1287/msom.2023.0644 .