Robotic Depalletizing: RFP Guide for Mixed Loads

The hardest part of robotic depalletizing is not lifting one neat, identical carton in a demonstration. Actual receiving areas see unregistered cartons, mixed SKUs, leaning loads, clear tape, moisture, crushed panels, loose labels, damaged pallets and gaps too narrow for a tool. A procurement team should therefore ask more than “What is the maximum speed?” The decisive question is: which loads can the cell accept, and how does it recover safely from everything else?

This guide is for manufacturers and logistics operations in Thailand and Southeast Asia preparing an RFP, FAT and SAT for unknown or mixed pallets. It complements our guides to picking robot implementation, palletizing robot implementation and robot gripper selection. Here, the entire receiving decision is evaluated as one chain: carton register → perception → grasping → extraction → exception recovery → safety. A supplier’s peak specification remains a reference, not a guaranteed result for your loads.

Define “unknown” before specifying a depalletizing robot

“Random,” “unknown carton” and “AI vision” can mean very different things. Dimensions may be unknown while appearance and weight are registered; several known SKUs may be mixed in one layer; or the incoming pallet may contain cartons never tested before. Bags, straps, corner boards and remaining stretch film may or may not be included.

The RFP should state the target level, excluded loads, expected frequency and required operating periods. Full automation of L4 is not automatically the best business design. Sending low-confidence cases to a controlled manual exception lane can improve overall availability and prevent damage.

Define rejection as a successful system outcome

Possible rejection conditions include excess weight, open tops, severe moisture, unstable stacks, remaining film, protruding metal straps, leaking contents or non-carton bags. If these are not defined, the FAT may show only good cartons and the SAT will expose repeated stops.

A controlled rejection is not a failure. If the cell identifies an unsafe load, reaches a safe state and tells the operator what to do, it has behaved correctly. Repeatedly attempting the same bad grasp until a carton tears or the stack collapses is unacceptable even if the nominal cycle is fast.

Step 1: Build a carton register that defines the acceptance envelope

Begin the RFP with a carton register, not a robot model. The register is a shared master for perception, grasping, transport, quality and safety.

Nominal values are insufficient. Board softens with humidity, tape moves between supplier lots and contents shift the centre of mass. Collect variation across season, supplier and shift. A condition proven in an air-conditioned test area may not remain valid at a humid or semi-open dock in Thailand.

Give each entry a status: approved, conditional, not evaluated or excluded. Define who approves a new carton, which sample and grasp tests are required, how vision is revalidated and how revisions are controlled. Otherwise every new SKU becomes an unplanned supplier service call.

Do not send only ideal samples for quotation

Provide boundary samples: minimum and maximum size, highest mass, weakest top panel, most reflective tape, low-texture surfaces, visually similar SKUs and cartons near the permitted crush limit. There is no universal responsible sample count; it depends on variation and consequence. The proposal should explain why its test set represents the actual population.

Step 2: Evaluate 3D vision depalletizing beyond the point cloud

A 3D sensor alone does not make a depalletizing solution. The system must locate the pallet, segment carton candidates, estimate faces and poses, propose collision-free grasps, decide removal order and transform results into robot coordinates. When confidence is insufficient, it must avoid speculative motion and move to rescanning or human review.

KUKA explains that its VisionTech can detect displaced carton positions in automatic palletizing and depalletizing. This is a useful vendor example, not a guarantee across every carton and lighting condition. Require recognition evidence using your own samples.

Common perception challenges include:

• black, glossy or clear-taped surfaces that cause missing or reflected depth;
• narrow gaps that make adjacent cartons appear as one object;
• crushed tops that no longer form a reliable plane;
• remaining film, straps, strings or document pouches;
• tilted pallets and changed reference surfaces;
• sunlight, vehicle lights, dust and dirty lenses;
• movement in a lower layer after the previous carton is removed.

During FAT, classify under-segmentation, over-segmentation, missed objects and wrong pose estimates. Do not reduce them to one average “recognition rate.” Report results by load class and by action: move, rescan, request confirmation or reject. A wrong positive that triggers motion may matter more than a safe non-detection.

Include calibration and rescanning in the cycle

The camera, robot base and pallet fixture must share a maintained coordinate relationship. Define recalibration after tool work, cleaning, impact or maintenance: responsible person, reference artifact, tolerance, record and release step. Break the time study into capture, inference, grasp planning, motion, grasp confirmation and discharge. Rescanning is not free, but it is often safer than forcing a grasp. Sustainable performance must include ordinary rescans and exceptions.

Step 3: Select the robot gripper for separation as well as holding

Robot gripper selection must prove that one carton can be separated without pulling its neighbours, held through the trajectory and released in the required orientation.

Use actual sealable area rather than total top area. Account for seams, labels, tape, recesses and holes. Monitor pressure, flow and changes during motion as appropriate to the risk. The required architecture and redundancy should follow the risk assessment, not a generic rule.

Robot payload is not carton mass alone. Include the tool, adapter, sensors and hoses, and check reach, pose, acceleration, centre of mass and inertia. Ask for worst-case load calculations, manufacturer selection evidence and the basis for speed limits.

ABB advertises one Robotic Depalletizer product with a maximum pallet height of 2.8 m, a peak of 750 cycles/h and payloads up to 30 kg. These are published nominal values for a specific offering. They are neither a TOMAS TECH guarantee nor proof for arbitrary mixed loads, environments and layouts. Your acceptance value must come from the agreed samples, layout, exception mix and downstream interface in FAT/SAT.

Prevent one extraction from destabilising the rest

The highest carton is not always the safest candidate. The tool may contact a neighbour; friction may pull a second carton; a lower support may be removed and leave another carton unstable. Grasp ranking should consider entry clearance, the swept volume of tool and arm, vertical extraction space and residual stack stability.

Step 4: Measure effective flow, not a peak cycle

A complete cycle includes image capture, segmentation, grasp planning, approach, grasp confirmation, extraction, transport, release and preparation. Pallet exchange, downstream blocking, carton reorientation, barcode reading and buffer recovery also affect output.

Measure separately:

• peak performance on easy accepted cartons;
• sustained performance with the representative SKU and defect mix;
• effective performance including rescans, re-grasps and human checks;
• shift performance including pallet exchange, cleaning and inspection;
• buffer and catch-up performance after downstream stops.

Use distributions, not only an average. Record median, slower-tail cases, downtime and failure classes. Targets must come from actual demand, takt, buffers and peak receiving data; this guide does not invent a universal cycle target.

This boundary table prevents responsibility from circulating during SAT. Define the owner and measurement method for every condition in the RFP.

Step 5: Design exception recovery as a normal function

Availability depends more on recovery than on the happy path. Typical cases include no detected carton, merged cartons, no safe grasp, failed vacuum, slip during lift, changed stack geometry, blocked discharge, barcode mismatch, entry into the safeguarded space, and camera/PLC/network faults.

For each exception, define automatic retry limits, rescan conditions, retreat position, quarantine destination, notification, access rights, logs and restart checks. Unlimited retry at the same grasp point can worsen damage and create a hazardous manual recovery.

The HMI should say what to do next

Operators need more than a module error code. Show which pallet and location failed, what was detected, the current safe state, and the approved next action. Time-synchronised snapshots, selected grasp, vacuum trace, previous motion and relevant sensor states help maintenance and suppliers work from the same evidence.

Manual intervention is part of the design. Specify access, lockout/tagout, stored energy, falling-load hazards and restoration of inventory state. Recovery speed must not be achieved by bypassing safety.

Step 6: Safety is more than adding a fence

ISO 10218-2:2025 was published in February 2025 and covers industrial robot applications and cells through design, integration, commissioning, operation, maintenance and decommissioning. Assessment concerns the integrated cell—robot, end effector, conveyors, safety-related controls, peripheral equipment and human tasks—not a robot model in isolation.

OSHA cautions that many robot incidents occur during non-routine work such as programming, maintenance, testing, setup and adjustment. In a depalletizing cell, this includes clearing crushed cartons, replacing suction cups, cleaning cameras, removing a load manually or editing a taught position. The SAT must test these interventions, not only automatic production.

Hazards can include collision and trapping, falling cartons or tool parts, unstable stacks, stored pneumatic/electrical/gravitational energy, combined conveyor movement, unexpected restart, sharp straps, damaged pallets, leakage and dust. Validate the adopted safeguards—doors, scanners, light curtains, emergency stops, restart prevention, safe speed or position—against the risk assessment and applicable requirements. Required performance levels cannot be prescribed responsibly without the cell-specific risk assessment. Measure real stopping time and distance, not only a PLC status bit.

Apply NIST’s task-driven performance approach

NIST presents performance assessment for robotic systems in task-driven terms spanning perception, mobility, dexterity and safety. For depalletizing, evaluate the task “safely remove one carton from this load and transfer it downstream,” rather than quoting sensor accuracy and robot repeatability separately.

Using the same tasks, samples and aggregation rules makes proposals more comparable. The buyer should receive the evaluation dataset for regression tests after new cartons or software changes.

What the RFP should require

Fix the scope, boundaries, evidence and responsibility without unnecessarily prescribing the solution.

Business and load requirements

• Carton register, pallet classes, arrival profile, shifts and required flow.
• Accepted, conditional, excluded and manual-bypass definitions.
• Boundary for film removal, pallet positioning and downstream transfer.
• Product damage, carton damage and wrong-routing criteria.

Technology and data requirements

• 3D camera arrangement, lighting, blind spots, cleaning and calibration.
• Confidence, rescan, grasp generation and removal-order behaviour.
• Gripper acceptance range, tool load and inspection or replacement conditions.
• PLC, WMS/MES/ERP, barcode and event-log interfaces.
• Ownership, retention and access for recipes, images, point clouds and logs.

Safety, operations and deliverables

• Risk assessment, applicable standards, safety requirements and validation plan.
• Non-routine work, lockout/tagout, manual recovery and training.
• Spares, consumables, diagnostics, remote support and cybersecurity.
• Layout, cycle breakdown, load calculations and tested-carton matrix.
• I/O, alarms, FAT/SAT protocols, maintenance standards, backups and restore procedures.

FAT: test limits and recovery, not only good cartons

FAT confirms the agreed design in the supplier environment. Separate what can be closed there from what remains for the site.

Each test record needs an ID, precondition, register and software revision, expected and actual result, evidence, deviation, corrective deadline and retest. Retain raw logs including failures. Calculate performance using the agreed load mix, not only supplier-selected ideal cartons.

SAT: prove the receiving decision in the Thai factory

SAT uses the real floor, light, humidity, power, air, network, upstream/downstream machines, workers and shifts. FAT approval does not guarantee SAT approval.

Confirm readiness of pallet positioning, film removal, conveyors, WMS data and work standards. Evaluate four timescales: single functional cases; sustained operation that reveals heat, contamination, leaks and queueing; shift operation including breaks, exchange and calls; and a period spanning supplier lots, weather and shifts. Duration and sample size must follow project risk. Use staged acceptance where short tests cannot close long-term uncertainty.

The final acceptance meeting should answer:

• Which register classes became automatic, conditional or manual?
• Where did perception, grasp, drop, damage or wrong-transfer events concentrate?
• Could the site team recover safely and repeat the procedure?
• Did effective output meet demand with upstream/downstream delays?
• Were safeguards and non-routine work validated and residual risks approved?
• Were software, recipes, logs, documents, training and spares delivered?
• Does every open item have an owner, date, retest and commercial consequence?

Make the investment decision on system value

Benefits may include less manual lifting, stable feeding, night operation, better data and consistent handling. Costs include equipment, safety, space, maintenance, consumables, power and air, carton onboarding, exception labour and backup staffing. A high automation percentage has little value if recovery is slow, damage rises or the downstream process cannot accept output.

Compare completed cartons, attended minutes, damage and misrouting, downtime, consumables, maintenance work, floor space and training against the baseline. A credible ROI requires project-specific quantities and costs; no universal payback is claimed here.

Thailand BOI reported 61 applications worth THB 7,071 million in Smart and Sustainable Industry in Q1 2026, including machinery upgrades and automation. This context shows continuing investment interest; it does not prove the return of an individual depalletizing project. Confirm eligibility and current conditions with official BOI channels.

Frequently asked questions

What is robotic depalletizing?

It uses robots, perception, grippers and material handling to remove loads from pallets and transfer them downstream. A production cell also includes removal order, drop prevention, exception recovery, safety and system interfaces.

Can a depalletizing robot handle mixed pallets?

Some solutions can, but “mixed” must be bounded. A load of registered SKUs is different from fully unknown cartons. Test the actual register, worst cases, confidence handling, graspability and exclusion logic in FAT/SAT.

What matters most in robot gripper selection?

Evaluate seal/contact area, carton strength, centre of mass, separation from neighbours, dynamic load, holding during loss or leakage, and consumable replacement—not maximum carton mass alone.

What 3D vision recognition rate is required?

There is no universal percentage. Set gates from the consequence of a wrong action, rescan time, human-review burden and actual load mix. Wrong positives and safe low-confidence behaviour matter alongside averages.

Can catalog cycles per hour be used as line capacity?

No. Cartons, pattern, rescans, grasping, travel, pallet exchange, exceptions and downstream waiting differ. Use catalog values for screening and prove sustained and shift performance with the agreed mix.

How are FAT and SAT different?

FAT validates design and limits mainly at the supplier; SAT validates the installed system with actual site conditions, interfaces and operators. Connect the same test IDs and close every FAT substitution at SAT.

Must the entire cell stop when an unknown carton arrives?

It depends on the hazard and process design. Some cases can move to a safe quarantine point, some can continue after human confirmation, and some require the cell to stop for removal. The essential requirements are to avoid forcing a low-confidence grasp and to make the state and recovery procedure explicit.

Conclusion: accept the whole chain from register to safety

Robotic depalletizing cannot be selected on speed alone. For unknown and mixed pallets, define the envelope in a carton register and evaluate 3D perception confidence, gripper separation and holding, effective flow, exception recovery and non-routine safety as one task. Specify evidence in the RFP, test limits and faults at FAT, and connect everything to actual Thai factory conditions at SAT.

TOMAS TECH can support carton-register preparation, site study, RFP requirements, vendor comparison and FAT/SAT planning. You can begin with load and exception scoping before committing to equipment. Contact TOMAS TECH.

References

• ISO 10218-2:2025
• OSHA Robotics
• ABB Robotic Depalletizer
• KUKA Automatic palletizing and depalletizing
• NIST Performance Assessment Framework for Robotic Systems
• Thailand BOI Q1 2026 Investment Applications