Integrated Optimization of Automated Warehouse Operations and Last-Mile Transport for Differentiated On-Demand Delivery
cs.LG
Quick verdict
Dual-DRL joint warehouse + last-mile optimization reports large simulated gains, but weak FIFO baselines, small scale, and no statistics temper the claims.
Summary
The paper proposes a bidirectional deep reinforcement learning framework that jointly optimizes AGV-based smart warehouse operations and last-mile delivery by a heterogeneous fleet. Inside the warehouse, MORM-AGDQN combines four reward machines (energy, inventory, external influence, safety), exponential reward shaping over Pareto front indices, CNN state encoding, and A*-guided epsilon-greedy action selection. Outside, MRMH-HCVRP uses a transformer encoder with vehicle-selection and node-selection decoders and a multiplicative reward over priority, travel-time, and congestion components. The systems exchange information: external traffic and customer priorities shape warehouse rewards, while the optimized order sequence feeds fleet routing. Using an Amazon delivery Kaggle dataset in a Python discrete-event simulation across baseline, peak-season, and uncertainty scenarios, the authors report 100% on-time warehouse departure/pickup, 97.06% last-mile on-time rate, 29.3-53.2% shorter average delivery time, 46.4% less travel distance (209 vs 390 km vs FIFO), and 92.31% high-priority satisfaction. Main limitations: baselines are limited to FIFO, historical data, and basic routing heuristics, with no comparison against commercial solvers or recent DRL routers; experiments are small-scale (up to 4 AGVs, 2-3 vehicles, 34-120 orders); results are single-run point estimates without variance or significance testing; internal inconsistencies exist (abstract's over-92% claim vs the 78.11% joint high-priority rate in Table 4); and no code is released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A bidirectional DRL framework synchronizing in-warehouse AGV scheduling with heterogeneous last-mile fleet routing outperforms independent and traditional optimization across warehouse, transport, and joint KPIs.
- MORM-AGDQN converges more smoothly and to higher cumulative rewards than PG, DQN, AGDQN, and MO-AGDQN, and generalizes across order scales (17 to 96+ orders).
- Joint optimization reduces last-mile average delivery time by 29.3-53.2% and total travel distance by 46.4% versus FIFO while achieving 100% on-time warehouse departure and inventory pickup.
- MRMH-HCVRP achieves the best cost-service balance among four routing methods (272.67 km, 3.41 h, 92% high-priority rapid delivery rate).
Strengths
Genuinely integrated scope
Section 4.3 and Figure 4 define a bidirectional coupling: Pareto-derived picking sequences become inputs to fleet routing, and external traffic/priority signals enter warehouse state (Eq. 36). The authors' own review (Section 2.3) confirms most prior joint-optimization work stops at in-warehouse processes or uses static MIP/heuristics, so the DRL-based end-to-end coupling is a real gap being addressed.
Structured, interpretable multi-objective reward design
Four separate reward machines (RMEn, RMIn, RMEx, RMSa, Section 4.2.1d) modularize objectives, and the exponential priority reward (Eqs. 45-48) derives base b and constant K in closed form from boundary rewards Rhigh/Rlow on Pareto front indices, making trade-off tuning explicit and auditable rather than ad hoc.
Algorithm-level ablation and generalization checks
Multi-perspective KPI reporting exposes trade-offs
Weaknesses
Weak baselines undermine headline gains
No statistical rigor
Small experimental scale contradicts scalability claims
Questionable multiplicative reward combination (Eq. 60)
Internal inconsistencies and editorial quality
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper on integrated warehouse and last-mile optimization achieve?
It proposes a deep reinforcement learning framework that jointly optimizes AGV operations inside a smart warehouse and delivery routing by a heterogeneous vehicle fleet outside. In simulation on Amazon delivery data, it reports 100% on-time warehouse departure, 29.3-53.2% shorter average delivery time, 46.4% less travel distance, and 92.31% high-priority satisfaction compared with FIFO ordering, though the comparison baselines are relatively weak.
What algorithms does the paper introduce?
Two: MORM-AGDQN, a multi-objective Deep Q-Network for warehouse AGVs that uses multiple reward machines, exponential reward shaping over Pareto priority fronts, and A*-guided action selection; and MRMH-HCVRP, a multi-head attention (transformer) network with separate vehicle-selection and customer-selection decoders for heterogeneous capacitated vehicle routing.
Is the method validated on real-world data or deployed in practice?
No. Validation is simulation-only, using a public Kaggle dataset of Amazon last-mile deliveries with Python discrete-event simulation and Graphhopper-generated routes, across three synthetic scenarios (normal, peak season with 2-5x order surges, and stochastic disruptions). No physical deployment or field experiment is reported.
Is code or data available for this paper?
The paper does not mention any code release. The underlying delivery dataset is public (a Kaggle Amazon delivery dataset), but simulation parameters, reward weights, and hyperparameter settings are not fully specified, making independent reproduction difficult.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All 15 figures could not be inspected from extracted text; convergence, Pareto-front, and routing-map claims rest on captions and prose descriptions only.
- Equations contain OCR artifacts and broken citations ('?') in the source text, so some notation judgments (e.g., Eq. 25) may partly reflect extraction noise rather than author error.
- Appendix pseudocode is partially garbled in extraction; hyperparameter details (learning rates, reward weights, penalty coefficients) may exist in the original but were not fully visible.
- This is an arXiv preprint (2609.19048); no peer-review verdict exists yet, so the review reflects the manuscript as submitted.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.