Warehouse slotting & picking optimisation

Order picking is typically the most labour-intensive warehouse process, and most of a picker's time is travel, not picking. Slotting (where each SKU lives) and picking logic (batching, zoning, waves, routing) decide how far people and robots walk. Combining demand forecasts with solvers and heuristics, and increasingly learned policies, cuts travel distance and raises lines per hour, while humans keep control of rules and exceptions.

Problem & value

• Today: many warehouses slot by arrival order or by an ABC list refreshed once a year. Pickers walk long paths, fast movers sit far from dispatch, and replenishment collides with picking. Wave planning is often done by experienced planners in spreadsheets or fixed WMS rules.
• AI value: (1) slotting by velocity, affinity (items ordered together), size and ergonomics; (2) order batching and wave/zone assignment by due time and carrier cut-off; (3) pick-path sequencing (a travelling-salesman variant on the aisle graph); (4) what-if simulation before moving stock; (5) dynamic re-slotting when demand shifts (seasons, promotions, new SKUs).
• Evidence: the academic literature consistently reports that order picking dominates warehouse operating cost and that travel is the largest time component; storage assignment, batching and routing each give double-digit travel reductions in published studies, but gains depend heavily on layout and order profile, so measure your own baseline rather than trusting generic percentages [1][2].

Variants

• E-commerce fulfilment: many small orders, high SKU count, strong affinity effects, peaks (Black Friday); batching and cluster picking matter most.
• Retail / wholesale distribution: case and layer picking, store-aisle sequencing, replenishment waves.
• Manufacturing intralogistics: kitting and line-side supply, with sequence constraints from OPT-002.
• Automated systems: AS/RS, shuttles, goods-to-person stations and AMR fleets (ROBO-002) where slotting is partly software-controlled (class-based storage in shuttles, order release sequencing).
• SMB manual warehouse: a one-off slotting study plus pick-path ordering inside the existing WMS.
• Cold chain / pharma: temperature zones, lot and expiry (FEFO), batch-segregation constraints.

How it works

WMS/ERP data (orders, stock, locations) + demand forecast (PRED-001)
  -> SKU profiling (velocity, cube-per-order index, affinity matrix, weight/ergonomics)
  -> slotting model (class-based or MIP/heuristic, re-slotting cost vs. benefit)
  -> order release: batching / zoning / wave building (solver + rules + due times)
  -> pick-path sequencing (shortest-route on aisle graph, congestion-aware)
  -> execution (RF scanner, pick-by-voice SPEECH-011, put wall, AMR dispatch ROBO-002)
  -> feedback (actual pick times, short picks) -> model recalibration

• Forecasting and affinity: gradient-boosted trees or time-series models give SKU demand; co-occurrence mining on order lines gives affinity clusters. Slow movers and seasonal items need separate handling.
• Optimisation core: slotting is an assignment problem (MIP or local search); batching is a bin-packing/clustering variant; path planning uses routing heuristics (S-shape, largest-gap, exact methods for single-block layouts, metaheuristics for complex ones). Solvers: constraint programming and MIP, plus ALNS-type metaheuristics.
• Learned policies: reinforcement learning for dynamic batching/order release and multi-robot assignment is productive in research and emerging in vendor products; it is trained in a digital-twin simulator and guarded by hard constraints.
• LLM role: LLMs do not replace the solver. They are useful as a planner copilot that explains plans, translates natural-language constraints into solver parameters and queries KPIs (see OPT-018).
• Human-in-the-loop: a planner approves the re-slotting move list (moves cost labour), sets business rules (hazmat, heavy items low, FEFO) and can override waves; floor supervisors handle exceptions.

Data preparation

Needed: item master with dimensions, weight and stackability, order lines per day (picking history), location master with capacities and zones, replenishment and movement logs, equipment constraints. In SMBs this lives in the WMS or ERP, often supplemented by Excel location lists and paper pick lists.

Typical gaps: missing or wrong item dimensions, locations that do not match the physical layout, bulk bookings without time stamps, no distinction between pick and replenishment movements, and manually overridden locations that were never booked.

Preparation steps: - Export 6-12 months of order lines and movements (PREP-005). - Measure and correct item master data for the A-movers first; this has the biggest effect. - Verify the location master against the real warehouse by a walk-through, and digitise paper maps or lists (PREP-008). - Tidy Excel layout files into tables (PREP-011). - Define metrics such as picks per hour and travel distance consistently.

Rough share of effort: about 30-40 percent, with master-data correction as main driver. GUIDE-003 gives a quick readiness check.

Implementation paths

Which path for whom

• SMB: Path A. Slotting and pick-path optimisation are available as modules of established WMS products, which fits standard warehouse processes without own development.
• Enterprise: Path B. Large or automated warehouses benefit from solvers fed with own WMS and order data, tuned to site-specific layouts and constraints.
• Developer: Path C. Open-source solvers (e.g. OR-Tools) run on HW1 and can be integrated into own WMS or logistics software.

Path A — Buy (recommended for most)

• WMS with built-in slotting/wave/path optimisation: Manhattan Associates, Blue Yonder (Luminate WMS), SAP Extended Warehouse Management (SAP EWM), Körber Supply Chain Software (HighJump heritage, DE/EU), Dematic and Honeywell Intelligrated / Zebra for automation software, SSI Schaefer WAMAS (DE/EU), Mecalux Easy WMS (ES/EU), TGW (AT/EU), Swisslog (CH, KUKA), Westernacher and Fiege/INTEGRA-style logistics software houses in DACH.
• Specialist optimisation tools: slotting/layout simulation add-ons (for example from WMS vendors or independent warehouse-design and simulation vendors), and AMR fleet managers with built-in task allocation (Locus Robotics, Geek+, Exotec (FR/EU), AutoStore ecosystem software (NO/EU)).
• Choose A when: you already run a WMS with an optimisation module, volumes are in the thousands of lines per day, and you want vendor-supported release and wave logic. Check: hosting region, API for slotting moves, support for your layout (multi-level, mezzanine), simulation capability and how the vendor handles tuning.

Path B — Build on cloud APIs

• Solvers as a service/library: Gurobi, FICO Xpress, HiGHS (open source) or Google OR-Tools CP-SAT, run in EU cloud regions; data in BigQuery/Snowflake/Azure Synapse EU regions.
• ML: managed AutoML or custom LightGBM/Prophet-style demand models on EU-region ML platforms.
• LLM layer (optional): a frontier model with tool calling (Claude, Gemini, GPT, Mistral) to let planners ask for what-if runs, with the solver exposed as a tool and outputs validated against schemas.
• Integration: WMS/ERP APIs (SAP EWM OData, Manhattan/Blue Yonder REST), orchestration with n8n, Airflow or Camunda.
• Choose B when: your layout or rules are non-standard, you need slotting for several sites with a shared model, or the WMS optimisation module is weak.

Path C — Self-host (open source)

• Solvers: Google OR-Tools (CP-SAT, routing library) [3], HiGHS, SCIP, OptaPlanner/Timefold (Java, open source; Timefold is EU-based, BE) for scheduling-style constraints, PuLP/Pyomo for modelling.
• Simulation: SimPy, AnyLogic (commercial), or custom discrete-event digital twin; Gymnasium/RLlib/Stable-Baselines3 for RL.
• Data/ML: Postgres + dbt, scikit-learn, LightGBM, statsforecast.
• Hardware: HW1 is sufficient: CPU server, with solver runs of seconds to minutes. A GPU only matters for large RL training (HW2).

Path D — Custom training

Worth it for large, stable operations (high-volume e-commerce, multi-robot goods-to-person) where a learned dispatching/batching policy beats rules. Needs a faithful simulator, months of order and task logs, and a fallback to rules. Otherwise tuned heuristics plus forecast are enough.

State of the art (2026)

• Hybrid approach dominates in production: forecast + MIP/heuristic for slotting, heuristic/metaheuristic for batching and routing, with simulation for validation. Pure end-to-end RL remains mostly research or confined to robot fleets.
• Learning-guided optimisation: neural models propose batches or routes, solvers repair and certify them; neural combinatorial optimisation shows promise on routing problems but is generally evaluated on benchmark instances, not on real warehouse constraints.
• Digital twins and simulation (including vendor tools and NVIDIA Omniverse-style warehouse twins) are used to test slotting and robot flows before moving stock.
• Affinity and ergonomics are first-class: golden-zone placement for heavy or fast items and workload balance.
• Robot-aware slotting for AMR and shuttle systems, where the optimisation target shifts to throughput and robot congestion.
• Limits and failure modes: poor master data (wrong dimensions, missing locations) is the commonest cause of failure; re-slotting costs labour and can exceed savings if done too often; forecasts fail for new SKUs and promotions; overfitting to last season's profile; optimised paths ignore real aisle congestion; planners distrust opaque plans, so explainability and override matter.

Shared resources & combinations

• Same solver stack (optimization-solver + simulation): OPT-001 vehicle routing, OPT-002 production scheduling, OPT-003 workforce scheduling, OPT-012 packing, OPT-018 natural-language optimisation copilot. One solver/simulation platform can serve all.
• Inputs from: PRED-001 demand forecast, OPT-005 replenishment and inventory policies, VIS-019 camera/drone stock counts for location accuracy.
• Execution with: ROBO-002 AMRs, ROBO-001 piece picking, SPEECH-011 voice picking.
• Downstream: OPT-001 outbound routing (dock and cut-off alignment), PRED-018 ETA.

Effort, cost & KPIs

• Path A: weeks to a few months for configuration and tuning inside a WMS; licence/subscription cost scales with sites and users.
• Path B: 3-6 months for a first production pipeline including WMS integration and simulation; costs are engineering time plus modest cloud spend.
• Path C: 3-6 months; mainly engineering and operations effort, little hardware.
• Cost drivers: data cleanup (master data and locations), integration with WMS, labour to execute re-slotting moves, change management on the floor.
• KPIs: lines (or units) picked per labour hour, walking distance per order line, order cycle time and on-time-in-full (OTIF), cut-off hit rate, replenishment moves per day, congestion/queue time, re-slotting cost vs. savings, picking accuracy.

People & change

Involve the warehouse manager as process owner and management for targets (throughput, cost per order line). The pickers are the real experts: invite a few of them to co-design slotting rules, since they know blocked aisles and heavy items. IT and the WMS vendor handle interfaces; the data owner of item master must release corrections.

Works council: depends, but often yes. Picking systems with scanner or time data are suited to monitor performance, so §87(1) no. 6 BetrVG applies as soon as pick rates per person are logged, which the Federal Labour Court treats as sufficient on objective suitability. §90(1) no. 3 requires early information on AI-planned work processes, and §80(3) allows an expert. Agree in a works agreement that optimisation data is evaluated per zone or process, not per person.

Personal data arises only through user IDs in pick logs: involve the DPO and aggregate or pseudonymise (PREP-022); a DPIA is rarely needed if no individual evaluation takes place. AI literacy: short briefing for pickers on why locations change (GUIDE-006). Pilot in one zone, communicate re-slotting dates early, and let supervisors override. Typical concern: "more pressure on picking rates". Address it openly (GUIDE-004, GUIDE-008).

Risks, regulation & compliance

• EU AI Act: typically minimal risk: optimising storage and routes is not an Annex III use. Caution: if the system allocates tasks to individual workers, monitors or evaluates their performance and behaviour (individual productivity scoring, automatic task assignment based on personal traits), Annex III point 4(b) (employment: task allocation, monitoring and evaluating performance) can make it high-risk. Prohibited practices (Art. 5) such as emotion recognition at the workplace must not be added to monitoring features. AI literacy (Art. 4) applies to staff operating the tools [5].
• GDPR and labour law: picker-level data (scanner timestamps, productivity rates) is personal employee data. Use aggregated or pseudonymised metrics, a clear purpose and legal basis, a DPIA where monitoring is intense, and involve the works council (in Germany co-determination under section 87(1) no. 6 BetrVG applies to technical systems capable of monitoring performance). Avoid surveillance-style pace pressure.
• Occupational safety: slotting should respect ergonomic limits (lifting height/weight), hazardous-goods segregation (ADR/GHS storage rules) and fire-protection limits; encode them as hard constraints.
• Data and security: WMS/ERP data is business-sensitive; secure solver services and APIs; for cloud, use EU regions and data processing agreements.
• Mitigations: planner approval of re-slotting, rule-based fallback, simulation before roll-out, constraint tests, audit log of plan versions, monitoring of fairness in task assignment across workers.

Getting started

1. Baseline: pick lines per hour, walking distance share, ABC/velocity profile, order-line distribution; clean SKU dimensions and location master data.
2. Run a slotting analysis (velocity, cube-per-order index, affinity) offline on 6-12 months of order lines and estimate travel savings with a simple simulation before any physical move.
3. Test the WMS vendor's built-in batching/wave/path functions against a solver prototype (OR-Tools) on the same historical orders.
4. Pilot in one zone with a limited move list; involve pickers and the works council from the start and keep worker-level metrics aggregated.
5. Review KPIs monthly and re-slot on a schedule justified by cost, not continuously.

Sources

1. de Koster, R., Le-Duc, T., Roodbergen, K.J. - Design and control of warehouse order picking: a literature review, European Journal of Operational Research 182 (2007) - https://doi.org/10.1016/j.ejor.2006.07.009
2. Boysen, N., de Koster, R., Weidinger, F. - Warehousing in the e-commerce era: a survey, European Journal of Operational Research 277 (2019) - https://doi.org/10.1016/j.ejor.2018.08.023
3. Google OR-Tools documentation (CP-SAT, routing) - https://developers.google.com/optimization
4. Timefold Solver (open-source constraint solver) - https://github.com/TimefoldAI/timefold-solver
5. Regulation (EU) 2024/1689 (AI Act), Art. 4, Art. 5 and Annex III - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. HiGHS open-source solver - https://highs.dev
7. SAP Extended Warehouse Management documentation - https://help.sap.com/docs/SAP_EXTENDED_WAREHOUSE_MANAGEMENT