Robotic truck & container unloading and mixed-case (de)palletising

Unloading floor-loaded trailers and sea containers is among the hardest, hottest and most injury-prone jobs in a warehouse. Mobile manipulators with 3D vision now segment the unstructured wall of mixed cartons, plan grasps and move boxes onto a conveyor; similar perception drives robots that build or break down mixed-SKU pallets. Operators use them to cover hard-to-staff dock shifts and to level inbound throughput, while people handle exceptions.

Problem & value

• Today: dock workers unload loose-loaded cartons by hand, often in heat or cold, with heavy repetitive lifting. Staffing is volatile, dock doors idle waiting for labour, and musculoskeletal injuries are a known occupational risk. Palletising mixed cases (store-ready or order-ready pallets) is done by hand or by rigid layer palletisers that only handle uniform SKUs.
• AI value: perception that handles unstructured box walls (varying size, colour, shrink-wrap, collapsed or tilted cartons) without a per-SKU teach-in, plus planners that compute stable mixed-pallet patterns.
• Reported figures: DHL Supply Chain states Boston Dynamics Stretch deployments reach up to 700 cases per hour, and signed a non-binding MoU in May 2025 for 1,000+ additional units by 2030 [1][2]. Treat vendor throughput claims as best-case; real rates depend on carton mix and trailer loading quality.

Variants

• Truck/trailer unloading (floor-loaded, mixed cartons) with a mobile base that advances into the trailer.
• Container unloading (20/40 ft sea containers, tighter geometry, heavier cartons, sometimes palletised goods).
• Truck loading (emerging, harder: stacking stability and load planning).
• Mixed-case palletising in distribution centres (store-order pallets; cases from AMRs or conveyors).
• Depalletising of mixed pallets into singulated case flow.
• Related but distinct: item picking from bins is ROBO-001; goods transport by AMRs is ROBO-002.

How it works

1. trailer/container docked
2. mobile base positions at the carton wall
3. RGB-D / 3D sensing
4. carton-wall segmentation (instance-level, edges, occlusions)
5. grasp selection (suction/vacuum, force limits, bottom-/side-grip)
6. motion plan
7. pick
8. place on conveyor / pallet (placement pose, collision check)
9. verification (drop detection, double-pick, dimension and weight check)
10. advance base / next layer
11. exception handling and remote assistance

• Perception: deep instance segmentation on depth and colour data finds individual cartons in a dense wall, including flush neighbours with no visible gap. Models are trained largely on synthetic plus real warehouse data (see ROBO-010).
• Planning: classical motion planning with learned grasp scoring; for palletising, 3D bin-packing solvers consider weight, fragility, support area and store layout (optimisation-solver).
• Control: high-speed vacuum grippers, force/torque and pressure sensing; learned policies are used for grasp choice, while the safety function remains deterministic.
• Human-in-the-loop: operators clear jammed or crushed cartons, supervise several units, and remote assistance teams may label a failed pick for the robot to retry (teleoperation-assist).

Data preparation

The key data are case dimensions and weights (item master in the ERP/WMS), ASN and order data, pallet patterns and rules (stacking, stability, customer requirements), plus camera images of real loads for tuning perception. They typically live in the WMS/ERP, in supplier EDI messages, in spreadsheets of the logistics planner and in the heads of experienced dock workers.

Typical gaps: missing or wrong dimensions and weights in the item master, duplicate or inconsistent article numbers, no master data for new or promotional packaging, pallet rules only as tribal knowledge, damaged or deformed cartons that never appear in the data, and no labelled images of difficult situations (shrink film, flat bags, tilted stacks).

Preparation steps: clean and complete the item master, ideally with automatic measuring at goods receipt (PREP-005, PREP-011); write down palletising rules in a structured form (PREP-016); define categories for exceptions (PREP-019); record and label images from your own docks over several weeks and build a test set (PREP-018, PREP-021). Vendors usually bring pretrained models, so own labelling is mainly for edge cases.

Rough share of effort: about 25 percent, mostly master data work in the WMS/ERP. GUIDE-003 gives a checklist.

Implementation paths

Which path for whom

• Enterprise: Path A, because mature unloading and mixed-case palletising robots from established vendors are proven in logistics and the dock process is standard.
• Developer: Path B, because integrators can combine cloud vision and planning APIs with robot hardware and warehouse systems when no product fits a special case.

Path A — Buy (recommended)

• Truck/container unloaders: Boston Dynamics Stretch (US; deployed by DHL in Europe) [1][3]; Pickle Robot (US); Dexterity Mech (US; two-arm mobile system, with a parcel-handling partnership); Anyware Robotics Pixmo (US); Mujin TruckBot (Japan, European presence); Honeywell Intelligrated and Siemens-aligned integrators offer unloading cells; EU players include Körber/KNAPP-class integrators for conveyor integration. Verify current EU availability, service coverage and CE conformity per vendor.
• Mixed-case palletising: Mujin MujinOS-based cells, Dexterity, Symbotic (US), Fizyr-based integrations (NL, vision software, EU), Pickit (BE, EU, 3D vision for robot cells), Photoneo / Brightpick-style 3D vision (SK/EU), plus robot makers ABB, FANUC, KUKA (DE) and Universal Robots with palletising packages. Robot-as-a-Service (RaaS) pricing is common.
• Choose A when: you have standard docks and carton mixes. Perception, safety certification and service make this a systems purchase, not a software project.

Path B — Build on cloud APIs

• No frontier-LLM API is a real-time controller. Use cloud for offline work: training/evaluating perception models, simulation (NVIDIA Isaac Sim/Omniverse on EU cloud regions), fleet telemetry, WMS integration and remote-assist tooling. Foundation models (VLMs) can help with labelling and exception triage, not with the pick loop.

Path C — Self-host (open source)

• Stack: ROS 2 with MoveIt 2 for planning, Open3D/PCL for point clouds, open segmentation models (e.g. Segment Anything family for labelling, YOLO/Mask R-CNN/Mask2Former-class for runtime), Isaac Lab or MuJoCo for simulation, OR-Tools for pallet planning. Open-weight robot policies (e.g. Octo, OpenVLA, pi0 family) are research-grade for this task.
• Hardware: edge GPU on the robot (Jetson Orin class, HW1) plus an industrial PC; training on HW3-HW4. Gripper, 3D camera, base and safety hardware are the real cost.
• Effort: a research-grade truck unloader is a multi-year engineering project; palletising cells are feasible for integrators.

Path D — Custom training

Worth it for a vendor or large operator with unusual cartons (polybags, sacks, tyres) or very high volumes. Needs thousands of annotated carton-wall scenes, real or synthetic, plus pick-success logs. Fleet data from deployed units is the main asset.

State of the art (2026)

• Production deployments exist at large 3PLs (e.g. DHL) and parcel networks; unloading is now a procurement category rather than a pilot, though fleet sizes remain small versus total dock labour [1][2].
• Trend from single-purpose arms to learned manipulation with a few robot-foundation-model elements (see ROBO-004) and expansion into case picking and truck loading.
• Limits and failure modes: collapsed or interlocked walls, bags and deformable packages, damaged or wet cartons, glossy shrink-wrap confusing depth sensors, heavy items (>20-25 kg) beyond gripper ratings, uneven trailer floors, double picks, and slow recovery after jams. Throughput varies strongly with load quality. Mixed pallets must also satisfy carrier stability rules.

Shared resources & combinations

• Same stack: 3D vision and grasp planning from ROBO-001; safety scanners and speed/separation logic from ROBO-016; simulation pipeline from ROBO-010; policy learning ideas from ROBO-003/ROBO-004.
• Combine with: ROBO-002 (AMRs move the pallets), OPT-006 (slotting and pick planning), VIS-019 (inventory counting), DOC-006 (ASN and logistics documents), ROBO-018 (exoskeletons for residual manual handling).

Effort, cost & KPIs

• Path A: 3-9 months from site survey to stable operation (dock changes, conveyors, WMS integration, safety validation); capex or RaaS fees per unit.
• Path C/D: years and a specialised team.
• Cost drivers: hardware, conveyor and dock modification, integration, safety validation, remote support.
• KPIs: cases per hour (sustained, not peak), pick success rate, interventions per 1,000 cases, uptime, damage rate, trailer turnaround time, injury rate and overtime at the dock, payback period.

People & change

Involve management, the logistics or warehouse lead (process owner), dock and palletising staff, IT (WMS/ERP interface), IT security, legal/compliance (machine safety, liability) and the works council from the start, ideally already in the discovery phase (GUIDE-001).

Works council: yes. Robots that take over unloading change work processes, and the systems typically log throughput, stops and operator interventions, which makes them objectively suited to monitor performance (§87(1) no. 6 BetrVG). The works council must also be informed in good time about AI-planned work (§90(1) no. 3) and may engage an expert (§80(3)). Public operators involve the Personalrat. A works agreement should exclude individual performance evaluation from robot data (GUIDE-004).

Data protection: cameras in docks capture employees and drivers; involve the DPO early, check purpose limitation, set retention periods and run a DPIA if monitoring of people is systematic (GUIDE-005).

AI literacy: support staff with training on how the robot behaves, what to do on faults and how to report errors (GUIDE-006). Main concern is job loss and physical strain relief versus replacement; communicate openly that the robot takes the heaviest work, offer retraining for roles such as robot operator or exception handler (GUIDE-008). Start with one dock as pilot group, define go/no-go KPIs (GUIDE-009) and keep humans in control: staff can stop the cell, handle exceptions manually and escalate issues.

Risks, regulation & compliance

• Machinery safety: the robot is a machine. Regulation (EU) 2023/1230 (Machinery Regulation) applies from 20 January 2027, replacing Directive 2006/42/EC; it adds requirements for machinery with self-evolving behaviour and AI-based safety components [4]. Robot safety follows ISO 10218 (industrial robots) and, for mobile operation, ISO 3691-4 and relevant ISO 13849 / IEC 62061 performance levels. Fencing is difficult in trailers, so safety scanners and exclusion zones are central.
• EU AI Act: an AI-based safety component of a machine falls under Annex I product legislation and can become high-risk (Art. 6(1)) with third-party conformity assessment; AI used purely for carton perception or productivity is usually minimal risk. Classification depends on whether the AI performs a safety function [5]. No Art. 5 prohibited practice is involved, but do not repurpose cameras for worker emotion inference or covert performance scoring.
• GDPR / works council: cameras and telemetry can capture dock staff; in Germany co-determination (BetrVG sec. 87) applies. Use privacy zones, no face recognition, and aggregate KPIs only.
• Workplace safety: risk assessment, lockout procedures, emergency stops, training for staff working near the robot.
• Operational risks: vendor lock-in, service dependence, cyber security of connected robots (NIS2, EU Cyber Resilience Act).

Getting started

1. Survey your docks: share of floor-loaded vs palletised loads, carton size/weight distribution, trailer types, current labour cost and injuries.
2. Collect video/depth samples of real carton walls, including the worst ones, and ask vendors for a proof on your freight.
3. Run a site acceptance test with sustained-rate, intervention and damage KPIs agreed in the contract.
4. Engage safety engineers early: risk assessment, CE conformity under the Machinery Regulation, works council.
5. Start with one dock door; plan conveyor and WMS integration before scaling.

Sources

1. DHL Group — MoU with Boston Dynamics for 1,000+ additional robots (13 May 2025) — https://group.dhl.com/en/media-relations/press-releases/2025/dhl-group-signs-mou-with-boston-dynamics-and-accelerates-cross-business-automation-strategy.html
2. The Robot Report — DHL buying 1,000+ Stretch robots from Boston Dynamics — https://www.therobotreport.com/dhl-buying-1000-stretch-robots-from-boston-dynamics/
3. Boston Dynamics — Stretch product page — https://bostondynamics.com/products/stretch/
4. Regulation (EU) 2023/1230 on machinery — https://eur-lex.europa.eu/eli/reg/2023/1230/oj
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. MoveIt 2 motion planning framework — https://moveit.ai/