Deliberate Assignment Deferral for Multi-Agent Pickup and Delivery with Deadlines

Automated warehouses must coordinate fleets of mobile robots online while meeting order deadlines. In online Multi-Agent Pickup and Delivery with Deadlines (MAPD-D), committing to a feasible task immediately may restrict flexibility and increase downstream tardiness through congestion and reservation interactions. We propose Deliberate Assignment Deferral (DAD-θ), a one-parameter gate on top of deadline-aware Token Passing baselines (D-TP and D-TPTS). At each token turn, the token holder evaluates the baseline-defined assignable tasks using the baseline score (lower is better); it commits only if the best score is at most θ, and otherwise follows the baseline fallback. A safety override forces assignment once any assignable task reaches non-positive pickup slack. We also introduce a scale-normalized score that makes θ dimensionless for transfer across maps and deadline scales. In 100-seed paired simulations across four arrival/deadline regimes on a benchmark map, scenario-calibrated DAD-θ reduces cumulative tardiness by 9–58% and increases on-time completion by 1.3–10.0 percentage points relative to always-assign. We discuss how θ can be calibrated offline in a digital twin and monitored online via deferral and safety-trigger rates for service-level control.