Load, container & cutting optimisation (packing/nesting)

Every shipment and every sheet of material is a puzzle: which items go where, in what orientation and order. Packing and nesting engines combine exact and heuristic solvers with learned policies to raise fill rates, cut empty-space shipping and reduce scrap, while respecting weight, stability, unloading-sequence and cutting-technology rules. Typical benefit: fewer trucks or containers and less raw material per output unit.

Problem & value

• Today: planners load by experience, spreadsheets or fixed pallet patterns; sheet-metal, wood, textile and glass shops nest parts with CAM tools or by hand. Results depend on the planner and on time pressure.
• Problem class: 3D bin packing, container loading, pallet building, 2D/1D cutting stock and irregular nesting are NP-hard. Real instances add practical constraints (axle loads, fragility, multi-drop sequencing, grain direction, guillotine cuts, remnant reuse).
• AI value: better utilisation of vehicle and container volume (fewer trips, lower CO2 per unit shipped), higher material yield and less scrap, fast re-planning when orders change, and automatic loading instructions for staff or robots.
• Evidence: vendors publish utilisation and yield gains, but figures depend heavily on the baseline and item mix; benchmark on your own orders rather than relying on marketing numbers. Academic work on learned online packing reports improvements over simple heuristics on synthetic data [3].

Variants

• Container and truck loading: 3D loading plans for sea containers, trailers and swap bodies, incl. multi-drop and LIFO sequencing and load-securing per EN 12195-1.
• Pallet building and mixed-case palletising: SKU-mix pallets for retail distribution and e-commerce; link to ROBO-014 for robotic execution.
• Parcel and carton selection ("cartonisation"): choose the smallest box for an order at pack-out; link to OPT-006.
• Cutting stock and nesting: 1D (bars, pipes, rolls), 2D (sheet metal, wood panels, glass, leather, textile marker making) and irregular nesting for laser, plasma, punching and CNC routing.
• Additive manufacturing build-volume packing: nesting parts in a 3D-printer build chamber.
• Private use: moving-house or car-boot packing apps are a niche; no dedicated products needed.

How it works

1. orders / part lists + item and container master data (WMS, ERP, CAD)
2. constraint model (dimensions, weight, orientation, stacking, sequence, cutting rules)
3. solver: constructive heuristic + local search / metaheuristic or exact CP-SAT / MIP
4. optional learned component (RL placement policy, predicted item mix, parameter tuning)
5. physics/stability check and simulation (centre of gravity, axle load, collision, kerf)
6. plan review by planner (3D viewer, manual edits)
7. instructions / CAM export / robot commands
8. actual-vs-plan feedback (damage, loading time, scrap) into parameter tuning

• Core: classical operations research does the heavy lifting: extreme-point and layer-building heuristics, genetic algorithms, simulated annealing, constraint programming (CP-SAT) and MIP for small or well-structured cases [1][2].
• Learning components: deep reinforcement learning trains placement policies for online packing, where boxes arrive one at a time without lookahead (robotic palletising, conveyor packing) [3]; ML models also predict dimensions or handling time from product master data and tune solver parameters.
• Vision: depth cameras (3d-vision) measure incoming parcels or remaining space for online and robotic packing.
• LLM role: minor; LLMs help translate planner requests into solver constraints (see OPT-018) or extract dimensions from packing lists (DOC-006).
• Human-in-the-loop: planners approve or edit plans; the driver or loader has final say on safety.

Data preparation

• Which data: item master data (dimensions, weight, stackability, orientation), orders and part lists, container, truck and sheet specifications, CAD files (DXF, STEP), historical loading and cutting jobs.
• Where it lives in an SMB: ERP or WMS item tables, Excel packing lists, supplier data sheets, CAD folders on file shares, packing lists as PDFs or paper.
• Typical gaps: missing or wrong dimensions and weights, no stackability or fragility flags, duplicate article numbers, inconsistent units, CAD files in mixed formats, packing lists only on paper.
• Steps: (1) pick the top 80 percent of volume and audit dimensions by measuring a sample (a dimensioning scanner helps); (2) add stackability, orientation and hazard attributes; (3) normalise units and remove duplicates, see PREP-011 for tidy tables; (4) export ERP or WMS data via PREP-005; (5) convert CAD to DXF; (6) read supplier packing lists with DOC-006. Tools: Excel or Power Query, Python/pandas, Paperless-ngx for paper lists.
• Share of effort: about 30-40 percent, and the main cost driver. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: A. Mature EU products for loading plans and nesting exist at SMB prices and the process is standard; machine makers often bundle a nesting package.
• Enterprise: B. Own packaging rules, special vehicles and tight WMS/TMS integration are best served by a solver service (OR-Tools, Hexaly, Gurobi) built on cloud APIs.
• Developer: C. OR-Tools CP-SAT and HiGHS run on a plain CPU server and give full control over constraints.

Path A — Buy (recommended for most SMBs)

• Load and container planning: Cargo-Planner (Sweden, EU), MixMove (Germany, EU, pallet and mixed-load planning), plus TMS/WMS modules from large vendors (SAP, Manhattan, Blue Yonder, Descartes) and TMS tools with built-in loading plans.
• Nesting and cutting: TRUMPF TruTops Boost (Germany, EU), Lantek (Spain, EU), Alma (France, EU), Lectra (France, EU, textiles/leather), HOMAG cutRite (Germany, EU, panel optimisation), Hypertherm ProNest, SigmaNEST and Radan; CAM and laser/plasma machine makers usually bundle a nesting package.
• Choose A when: standard item types and machines, and when integration into the existing WMS/CAM matters more than custom algorithms.
• Check: constraint coverage (axle load, LIFO, grain, remnants), import formats (DXF, STEP, EDI), API access, run time for typical orders, on-prem or EU-hosted option.

Path B — Build on cloud APIs

• Solver as a service or library calls: Gurobi, Hexaly (France, EU) and CPLEX for MIP/local-search models; Google OR-Tools CP-SAT as the free alternative [1][2].
• Orchestration: microservice receiving order lists, returning plans as JSON and 3D visualisation in a web viewer (three.js); results written back to WMS/TMS.
• LLM layer (optional): an LLM via an EU-hosted endpoint (Mistral, or hyperscaler EU regions) to parse free-text load requests or packing lists into constraint objects.
• Choose B when: you have unusual constraints (own packaging rules, special vehicles) that off-the-shelf tools do not model.

Path C — Self-host (open source)

• Solvers: Google OR-Tools (CP-SAT, bin-packing and knapsack models) [1][2], HiGHS and SCIP for MIP, py3dbp and similar Python 3D packing libraries for prototypes [4].
• Nesting: SVGnest and DeepNest for irregular 2D nesting, suitable for hobby or small-shop laser cutting [5][6]; for production use check licence and robustness.
• Learned policies: PyTorch with open RL environments for online 3D packing; Isaac Sim or PyBullet for stability simulation.
• Hardware: HW1 (CPU server); solvers are CPU-bound. A GPU (HW2) only helps for RL training and vision.
• Ops effort: needs an OR-literate developer; plan for model maintenance and test sets.

Path D — Custom training

Worthwhile for high-volume robotic palletising or conveyor packing with online arrivals, where a learned policy trained in simulation can beat heuristics, or for tuning solver parameters per customer instance class. Data: historical orders with item dimensions, and simulation environments; sim-to-real gap and stability under vibration need physical validation.

State of the art (2026)

• Hybrid approaches dominate: strong heuristics and CP-SAT or MIP give the baseline; learning adds guidance (which item next, which placement) rather than replacing the solver.
• Online 3D packing with RL is an active research field with simulation benchmarks, with growing but still limited industrial deployment, mainly in robotic palletising [3].
• Practical constraints (stability, load securing, unloading order, mixed-weight stacking) remain the gap between benchmark results and real loading plans; plans must be verified against EN 12195-1 and vehicle limits.
• Nesting has matured in CAM suites with automatic remnant handling and common-line cutting; AI mostly adds automation of job preparation (part recognition from CAD).
• Limits: declared item dimensions are often wrong; soft goods deform; planned layouts break when loaders deviate. Optimal plans are only useful if they can be executed quickly.

Shared resources & combinations

• Solver stack (optimization-solver): shared with OPT-001 (routing), OPT-002 (scheduling), OPT-006 (slotting) and OPT-008; one CP-SAT/MIP service and test-data pipeline serves all of them.
• Upstream: OPT-005 (replenishment) and OPT-001 (tour planning) determine what is loaded onto which vehicle.
• Downstream: ROBO-014 robotic (de)palletising, OPT-037 container terminal stacking.
• Related: OPT-009 (process parameters on cutting machines), PRED-032 (cost estimation from CAD, which uses nesting yield), DOC-006 (reading packing lists), OPT-033 (freight matching with load utilisation).

Effort, cost & KPIs

• Path A: days to weeks; subscription or licence, nesting suites typically licence-based.
• Path B: 2–4 months for a production service with WMS/TMS integration.
• Path C: 1–3 months for a prototype, longer for production hardening.
• Path D: 6+ months incl. simulation and robot integration.
• Cost drivers: master-data quality (item dimensions), constraint modelling, integration.
• KPIs: volume and weight utilisation per vehicle, trips or containers per period, material yield (% of sheet used), scrap and remnant reuse rate, planning time per order, loading time, damage claims, CO2 per tonne-km.

People & change

• Involve early: management (target KPIs), the logistics or production lead (process owner), planners and loaders (users, real constraint experts), IT (ERP/WMS/CAM interfaces) and the owners of item master data.
• Works council: depends. The optimiser works on goods, not people, so usually little monitoring concern. It becomes relevant when the tool logs which planner edited plans or measures loading times per person; then §87(1) no. 6 BetrVG applies. Inform the works council under §90(1) no. 3 when loading or planning processes change. Public sector: staff council.
• Data protection: item data is not personal; consignee names and addresses are, so send only dimensions to cloud services and sign an AVV (Art. 28 GDPR) if order data is shared. A DPO check is enough, no DPIA expected.
• AI literacy: support planners and loaders with a short introduction to how plans are generated and when to override (Art. 4 AI Act), see GUIDE-006.
• Pilot and communication: test 20-50 historical orders with the loading crew, then one route or machine. Typical concerns are unrealistic plans and loss of planner expertise; address them by adding constraints together with the crew, allowing manual edits and giving the driver or loader the final say on safety.
• Guides: GUIDE-008, GUIDE-009, GUIDE-004.

Risks, regulation & compliance

• EU AI Act: minimal risk. Packing and nesting are not Annex III use cases; AI literacy (Art. 4) applies. Robot cells that execute plans fall under machinery safety rules (Regulation (EU) 2023/1230), not the AI Act's high-risk categories unless the AI is a safety component [7].
• GDPR: item and shipment data is generally not personal; consignee names and addresses in order data are, so apply data minimisation when sending order data to cloud APIs (usually unnecessary: send only dimensions).
• Transport safety: load securing (EN 12195-1, national rules such as VDI 2700 series in Germany), axle-load and gross-weight limits remain driver and shipper responsibilities; software cannot transfer liability.
• Risks: wrong master data causing infeasible plans; unstable loads; vendor lock-in on proprietary nesting formats. Mitigations: dimension audits, safety margins, plan validation rules, export in open formats (DXF), human sign-off before release.

Getting started

1. Measure the baseline: average fill rate or material yield, trips/containers per month, planning effort.
2. Clean item master data (dimensions, weights, stackability) for the top 80% of volume.
3. Run 20–50 historical orders or cutting jobs through 2–3 tools (e.g. Cargo-Planner, OR-Tools prototype).
4. Add constraints one by one (LIFO, axle load, grain) and check plans with the loading crew.
5. Integrate with WMS/TMS or CAM and track utilisation and scrap monthly.

Sources

1. Google OR-Tools — Bin packing — https://developers.google.com/optimization/pack/bin_packing
2. Google OR-Tools — CP-SAT solver — https://developers.google.com/optimization/cp/cp_solver
3. Zhao et al. — Online 3D Bin Packing with Constrained Deep Reinforcement Learning (AAAI 2021) — https://arxiv.org/abs/2006.14978
4. py3dbp — 3D bin packing library — https://github.com/enzoruiz/3dbinpacking
5. SVGnest — https://github.com/Jack000/SVGnest
6. DeepNest — https://github.com/Jack000/Deepnest
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Regulation (EU) 2023/1230 (Machinery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1230/oj