AI Can Find the Pallet. Should Your Heavy-Payload AMR Pick It Up
The next buying question is whether the robot should commit
AI-enabled pallet handling is becoming a concrete purchasing option. In its June 2026 Automate announcement, Teradyne Robotics described the MiR1200 Pallet Jack as commercially available and outlined a demonstration connecting mobile robots with a palletizing application. That announcement establishes a product and integration direction, not a measured pickup success rate for a buyer's warehouse. [1] The practical question for procurement is now more demanding: when a robot recognizes a pallet, what evidence allows it to begin the pickup?
AI pallet detection identifies a candidate target. A production-ready application must additionally establish that the target belongs to the mission, matches an approved carrier configuration, can be approached and engaged, and remains suitable for the intended movement. These are separate claims. A camera can locate a pallet accurately while the application should still refuse to collect it.
This analysis concerns floor-level pickup by autonomous pallet jacks and fork-equipped mobile robots handling substantial unit loads. It does not assume that every AMR chassis has forks, that every pallet construction is compatible, or that floor-level evidence qualifies an application for high-rack handling. The scenarios, decision model and numerical example below are original engineering illustrations, not results from a named supplier or an operating installation.
The purchasing objective is useful discrimination. The system should accommodate approved variation without demanding unnecessary manual alignment, while identifying conditions that require another observation, a corrected presentation or human assessment. Maximizing recognition alone can reward the wrong behavior. Refusing every difficult case can also produce an apparently cautious system that delivers little operational value.
Three pallets that look similar but require different decisions
Presentation A: the pallet is skewed, but the pickup remains feasible
Imagine an approved pallet placed diagonally in a collection bay. Its load is secured, its required entry openings are unobstructed, and its actual position remains inside the application's validated approach envelope. The target is imperfectly presented, but the imperfection is one the system was purchased to handle.
Here, pallet pose estimation should support a revised approach based on the observed carrier, rather than a blind move toward the bay's nominal coordinates. The robot still needs enough maneuvering space, valid measurements and an approved engagement path. A high-confidence pose does not create extra turning room or make an unsuitable entry side compatible with the forks.
The commercial benefit should be measured as reduced manual repositioning for this approved population. A demonstration that succeeds only after a technician straightens each pallet does not establish that benefit. Record who changed the presentation, when it changed and whether the subsequent pickup used fresh observations.
Presentation B: the carrier is recognizable, but the entry path is obstructed
The second pallet has familiar dimensions and enough visible structure to locate its face. Loose packaging film crosses an opening, or a displaced board interferes with the intended entry path. Recognition may remain successful because the algorithm can infer geometry from other visible features.
The distinction is physical: inferring where an opening belongs is different from proving that the required volume is available. The current ifm getPallet documentation describes configurable pallet geometry and detection approaches with different capabilities under partial occlusion. That is evidence about locating the target; it is not permission to drive through an obstruction. [2]
For this case, the acceptance requirement should specify an appropriate hold or escalation when entry clearance cannot be established. Repeatedly observing the same obstructed opening may improve the estimate of its nominal center while doing nothing to remove the film. A retry is valuable only when it can resolve the uncertainty relevant to the next action.
Presentation C: the openings are clear, but the carrier is unacceptable
The third pallet presents clean openings and a stable-looking outline. However, a support block is broken, the load restraint is visibly compromised, or the carrier belongs to a construction that the application has never approved. Its geometric appearance may make it easier to detect than either of the first two examples.
Useful damaged pallet detection therefore requires a defined inspection scope. Which defects are observable from the installed viewpoint? Which require inspection elsewhere? Which conditions are outside the system's capability? A missing visible component and an internal weakness hidden beneath the load should not be represented as equally inspectable.
SICK's March 2025 PAIS announcement describes a dedicated 3D pallet integrity inspection application for structural defects. Its existence illustrates the distinction between locating a pallet and inspecting its condition. It does not establish that an ordinary mobile-robot perception camera performs an equivalent inspection, or that any visual inspection guarantees remaining mechanical strength. [4]
Write a pickup eligibility specification before requesting an accuracy figure
A pallet recognition system needs an operational definition of an acceptable input. Procurement should develop that definition with warehouse operations, packaging or carrier specialists, the vehicle integrator and the team responsible for the transported product. Leaving it until acceptance testing creates a predictable dispute: the supplier calls a refusal correct while operations calls the same pallet normal production.
Start with a carrier family, not merely a nominal outside dimension. Record construction, intended entry face, pocket geometry, underside features and any equipment-specific restrictions. Two carriers can share a footprint while presenting very different engagement conditions. Include empty and loaded configurations separately when their observable features or handling requirements differ.
Then define the combined handling case: carrier and load mass, approved load distribution, load envelope, restraint condition and permitted overhang. These limits must come from the equipment documentation and application engineering. This article supplies no universal mass, tilt, clearance or damage threshold because an invented generic limit would obscure the actual design boundary.
| Decision area | Information the project should define | Example of an unresolved case |
|---|---|---|
| Carrier compatibility | Construction, entry side, geometry and approved variants | A familiar footprint with an unqualified underside |
| Mission identity | How the physical load is matched to the requested movement | Two similar pallets occupying adjacent positions |
| Presentation | Allowed position, orientation, support surface and access space | A measurable pallet outside the feasible approach envelope |
| Condition | Observable defects, upstream inspection and escalation rules | A hidden support area that the camera cannot inspect |
| Load configuration | Mass, distribution, restraint and outer envelope | A correctly identified carrier with an unapproved load |
The output should distinguish an approved case, a known unacceptable case and an unresolved case. Unknown should remain visible in the data. Forcing uncertainty into either a positive classification or an undifferentiated failure code removes information that engineering and operations need to improve the process.
Product scope also matters. MiR's current MiR1200 product page specifies EU-pallet handling and describes AI-based perception. Buyers should read such statements together with the applicable configuration and supplier documentation, rather than extending them to every pallet format found in a global supply chain. [7] Compatibility is a purchasing question before it becomes a software tuning question.
Follow the evidence from pixels to a supported load
Finding an object does not establish its business identity
A detection can establish that a pallet-like object is present without establishing which inventory unit it carries. If the mission requests a specific shipment, the application needs a reliable association between that physical presentation and the business record. Depending on the installation, that may involve a controlled bay assignment, a scanned identifier or another validated identification process.
Challenge the association when neighboring pallets look alike, when a worker exchanges positions, and when a previously assigned load is removed before arrival. A correct pickup of the wrong order remains a failed logistics transaction. Avoid crediting the perception algorithm with identity information that actually comes from an upstream process and may have become stale.
Measured geometry must retain its meaning through the controller
3D pallet detection can provide useful spatial information, but the receiving software must know its coordinate system, capture time, units and validity. SICK's 2023 Pallet Pocket Detection manual describes output in device coordinates for a vehicle controller to use. The documented example is a manned-forklift application; its relevance here is the separation between a sensor measurement and a vehicle decision, not proof of autonomous performance. [3]
One subtle integration risk is treating an available field as a measured quantity. In the current ifm getPallet documentation, pitch estimation is disabled by default and the corresponding rotY result is zero. A downstream application must not interpret that default as evidence that the physical pallet is level. [2] Similar questions belong in interface reviews for every value that can be unavailable, substituted or retained from an earlier observation.
The transform between camera, vehicle and forks must also remain valid after installation and maintenance. Robot localization in a warehouse map does not, by itself, establish the relative position of the pallet openings. Conversely, an excellent local observation cannot repair an invalid vehicle configuration. The separate guide to AMR docking tolerance verification examines the geometric chain in greater detail.
Recognition confidence needs an explicit interpretation
A score from automated forklift vision may describe a model's detection output, matching quality or another supplier-defined quantity. Unless validated for a specific interpretation, it should not be presented as the probability that a complete pickup will succeed. The event being predicted must be stated before a percentage has operational meaning.
Guo and colleagues' ICML 2017 paper treats confidence calibration as the relationship between predicted confidence and observed correctness. It is relevant methodological background, not a pallet-handling benchmark. [5] For a procurement trial, request evidence from the delivered model and the buyer's operating population rather than borrowing a calibration claim from an unrelated dataset.
Even a well-calibrated recognition score does not certify personnel protection, carrier strength or adequate load support. Those claims require their own evidence and implemented functions. The site's discussion of safety-rated sensing and navigation perception explains why ordinary perception outputs should not silently replace a credited protective function.
Engagement and load support require additional observations
Pallet pocket detection addresses a target needed for entry. A complete application must also determine whether the intended engagement occurred and whether the subsequent load state satisfies its approved handling conditions. A successful command transmission is not a physical observation, and a visible carrier is not necessarily supported as intended.
The ifm volCheck documentation illustrates a separate observation capability: checking a defined three-dimensional volume and reporting valid-pixel information and distance. Listed applications include load presence and space occupancy. Those outputs can contribute evidence within a designed system; the documentation does not turn them into a universal proof of correct support or a personnel-protection certification. [6]
Define completion using the observations available in the installed application. Keep the physical result and the business transaction consistent, including exceptions after partial progress. The related guide to AMR transfer handshake validation covers the broader problem of distinguishing a command, an acknowledgment and a confirmed transfer.
A validation dataset should represent decisions, not impressive frame counts
Useful pallet detection validation starts with a case definition. For this application, a case can be one identified pallet presentation for one requested pickup, with its initial condition, permitted observations, decision and eventual outcome retained together. A camera producing hundreds of frames during that case has not created hundreds of independent production opportunities.
Keep all attempts under the original case identifier. Otherwise, a difficult pallet that generates several failed observations followed by a successful one can inflate both the sample count and the apparent success rate. Retain timeouts, unresolved cases and interventions. Define in advance how a changed pallet, replaced load or materially altered presentation starts a genuinely new case. Link any new record to the original mission, retain the earlier held or failed outcome, and preserve the history of manual intervention.
Separate ordinary production sampling from deliberate challenges
One dataset should describe the incoming population the operation expects to handle. Sample across relevant suppliers, carrier ages, load types, shifts and collection points. Another dataset should deliberately test known boundaries and difficult combinations. Both are useful, but they answer different questions: expected operating performance and behavior under specified challenges.
A challenge set containing many damaged or unsupported carriers must not be used directly to forecast daily exception labor. Likewise, a routine sample with almost no difficult presentations cannot establish behavior at the operating boundary. Report the mix and results by meaningful subgroup before presenting an aggregate number.
Use the existing AMR environmental edge-case testing framework for illumination and contamination mechanisms. For this pickup study, connect each selected condition to a decision: identity unresolved, geometry unavailable, entry clearance uncertain or approved pickup completed. This keeps the test focused on consequences rather than accumulating visually dramatic photographs.
Prevent the evaluation from rehearsing the training set
Reserve evaluation cases that were not used to tune the delivered configuration. Splitting adjacent video frames randomly between development and evaluation leaves nearly identical scenes on both sides. Where practical, separate physical pallets, acquisition sessions, supplier batches or sites according to the generalization claim the supplier is making.
Ground truth also needs governance. Qualified reviewers should apply the agreed eligibility specification, document disagreements and preserve unresolved labels for adjudication. If a defect cannot be verified from an image, an image-only reviewer should not invent certainty. Physical inspection or another independent record may be necessary to establish what the system was expected to decide.
Use recorded data, suitable test objects and motion-inhibited evaluation for inappropriate pickup candidates. Potentially unsafe loaded pallets should not be lifted merely to demonstrate that a perception gate can reject them. Physical execution trials require their own approved conditions and controls; passing an offline classifier test is not automatic authorization to move.
One headline accuracy figure can hide three different purchasing outcomes
Consider the following synthetic example. The numbers are created solely to show how denominators change the interpretation; they are not supplier results, an acceptance standard or a forecast. A deliberately selected evaluation contains 1,000 object-level presentations: 900 eligible and 100 ineligible under the previously agreed rules. Ineligibility includes incompatible carriers and incorrect mission identity as well as unacceptable physical conditions.
| Independently assessed presentation | Gate would allow pickup | Gate would hold pickup | Total |
|---|---|---|---|
| Eligible | 864 | 36 | 900 |
| Ineligible | 4 | 96 | 100 |
| Total | 868 | 132 | 1,000 |
Overall decision accuracy is 96%: 864 correct approvals plus 96 correct holds, divided by 1,000. Yet the gate would incorrectly admit four of the 100 ineligible presentations, an invalid-admission rate of 4%. It would also hold 36 eligible presentations, admitting 96% of the eligible population. These consequences need separate review; an aggregate accuracy figure cannot decide whether either result is acceptable.
The four incorrect approvals are discovered in the offline or motion-inhibited stage. They are not executed pickups and must not be described as observed mechanical failures. Investigate their causes and the applicable release criteria before authorizing physical operation. This illustrative dataset establishes neither an acceptable risk level nor a safety failure probability.
Now consider an illustrative physical trial of the 864 eligible presentations that the gate admitted. Their relevant conditions are maintained or re-established and verified before each authorized trial. Of these, 840 complete on the first attempt. Another 16 complete after one approved, fully automatic re-observation or repositioning before contact. Eight remain incomplete. Retries remain attached to the original cases.
| Measure | Calculation | Result |
|---|---|---|
| First-pass completion among admitted eligible cases | 840 / 864 | 97.22% |
| First-pass completion across all eligible presentations | 840 / 900 | 93.33% |
| Automatic completion across all eligible presentations, including permitted retry | (840 + 16) / 900 | 95.11% |
| Eligible cases remaining for human review | 36 held + 8 incomplete | 44 cases |
For autonomous pallet pickup, the third measure describes useful completion across the eligible workload more directly than the first. The first remains valuable for understanding performance after admission. Neither should replace the other, and the eight incomplete cases must remain in the denominator even if a later investigation attributes them to something other than vision.
If each of the 44 eligible exceptions takes an assumed five minutes of human review, that is 220 person-minutes of review. It is not a promise that each problem is resolved in five minutes. The calculation excludes handling of the ineligible population, investigation of the four incorrect approvals, travel, queueing, routine support and maintenance.
Because the example uses a deliberately selected 900-to-100 mix, neither its aggregate accuracy nor its labor total should be projected onto a real shift. Forecasting requires the actual prevalence of each presentation type and measured intervention times. A change in incoming packaging or pallet supply can change the operational result even if the model's performance within each subgroup stays the same.
Recovery should resolve a specific uncertainty
Before contact, another observation may be useful
A pre-contact retry should have a reason: obtain a less obstructed view, refresh a stale observation or use an approved approach position that reveals the required geometry. Define the permitted actions, elapsed-time limit and stopping condition in the application design. There is no universal retry count that suits every load, bay and vehicle.
The controller should not silently reduce its eligibility requirements because the first attempt failed. A different observation can supply missing evidence; weakening the evidence requirement changes the application. Record which uncertainty the retry was intended to resolve and whether the resulting information actually changed the decision.
After contact, the problem has changed
Once a fork or load-handling element may have engaged the carrier, a repeat of the original approach is no longer merely another vision attempt. The load may have moved, support may be partial, and the previous scene may no longer describe the physical state. Recovery needs the equipment-specific, validated handling procedure appropriate to that state.
A remote acknowledgment should not be interpreted as proof that the carrier is correctly supported or the space is clear. The guide to AMR safe restart and recovery develops that distinction. In the pickup record, retain the last verified physical state and the evidence required before another movement is permitted.
Manual assistance belongs in the performance record
If an operator removes film, repairs a presentation, changes a carrier or selects the correct inventory record, record that work as an intervention. A successful subsequent pickup does not retroactively become an unattended first-pass success. Conversely, a correct refusal of an unacceptable carrier should not be mislabeled as a recognition failure.
Assign ownership for each exception category. Packaging problems may belong with receiving or the supplying plant; unavailable geometry may require the integrator; identity mismatches may require inventory control. Operations should not have to diagnose every exception from a generic red indicator. A useful message identifies the unresolved condition, relevant evidence and the responsible workflow.
Turn the pilot into a purchasing agreement that survives production changes
For AMR pallet handling, acceptance should combine correct decisions with useful completed work. State the eligible population, excluded configurations, required handling of unresolved cases, permitted recovery actions and measurement rules. Report incorrect approvals separately from eligible-load refusals, and report completed movements separately from recognition events.
Agree on sample coverage and uncertainty before the pilot starts. A run with no observed incorrect approvals does not establish that their true probability is zero, especially when many cases reuse the same physical carriers under similar conditions. The acceptance team should review the number and diversity of cases alongside the observed rates.
Measure elapsed time at the task level, including observation, permitted retries and any queue effects relevant to the operation. A model's inference time is only one contributor. A configuration that improves recognition but causes repeated repositioning can reduce productive throughput. The commercial comparison should include intervention workload, blocked collection space and downstream disruption under the same workload assumptions.
Require an evidence export sufficient to reconstruct exceptions: case and mission identifiers, relevant observation timestamps, carrier configuration, software and model versions, eligibility reason, attempt history and final physical outcome. Agree what imagery or depth data will be retained, for how long, and who can access it. The purpose is reproducible diagnosis, not indiscriminate collection.
Finally, define change control. A new pallet supplier, packaging film, camera mount, model version or handling attachment can alter the validated conditions. Preserve representative ordinary cases and important counterexamples for regression evaluation. After a change, compare both incorrect admissions and unnecessary refusals; improving one metric does not guarantee that the other improved.
The deliverable is a repeatable operating boundary with an owner. Receiving teams know which presentations are supported, engineers know which evidence authorizes progression, and procurement can distinguish a capability gap from an upstream process defect. That agreement is what converts a convincing vision demonstration into a manageable material-handling service.
Focused FAQ
Does recognizing a damaged pallet mean the robot can transport it?
No. Recognition means the system identified relevant visual or geometric features. Transport eligibility additionally depends on the approved carrier condition, load configuration and handling requirements. A system may correctly recognize a damaged pallet and correctly refuse pickup. Hidden defects may require a separate inspection process.
Can one detection setup handle every pallet construction?
That should not be assumed. Require the supplier to identify supported constructions, entry faces and configuration limits, then test the actual incoming population. Matching outside dimensions alone does not establish compatible pocket geometry, underside clearance or load support. Unqualified variants should remain explicit exceptions until evaluated.
What is the difference between location accuracy and pickup success?
Location accuracy measures the error in a defined position or orientation estimate. Pickup success measures an agreed physical task outcome. An accurate estimate can accompany a refused pickup, an obstructed entry or the wrong inventory target. Report both metrics with their own units, denominators and eligibility conditions.
Should every refusal count against the supplier?
No. A correct refusal protects the agreed operating boundary. However, repeated refusals of eligible production loads reduce usable capacity and should be measured. Acceptance needs both sides: whether inappropriate candidates are held and whether approved work completes with the agreed level of intervention.
How should automatic retries be counted?
Keep them within the original presentation record. Report first-pass completion, eventual automatic completion within the approved retry policy, elapsed time and manual intervention separately. A worker correcting the presentation creates assisted recovery; it should not be hidden inside an autonomous completion claim.
What should a buyer request before expanding to another warehouse?
Request a comparison of pallet families, packaging, collection geometry, operating conditions and exception ownership between sites. Evaluate representative local cases using the intended configuration. Transfer of software does not, by itself, transfer evidence for a different incoming population or physical installation.
Sources and evidence notes
Sources checked on September 29, 2026. Manufacturer publications establish documented features or stated application scope; they are not independent performance certification. The three presentation scenarios, eligibility framework and numerical dataset are original analytical examples.
- Universal Robots: Teradyne Robotics Automate 2026 announcement. Published June 18, 2026, with a June 11 dateline. Supports product availability and the announced demonstration scope.
- ifm O3R documentation: getPallet. Current documentation consulted at version 1.21.26. Supports the limited statements about geometry configuration, occlusion and the default pitch output.
- SICK: Pallet Pocket Detection operating information. Edition dated May 17, 2023. Supports the device-coordinate output and controller interface distinction.
- SICK: PAIS pallet integrity inspection announcement. March 26, 2025. Supports the distinction between dedicated condition inspection and target localization.
- Guo et al.: On Calibration of Modern Neural Networks. ICML 2017, PMLR 70, pages 1321–1330. Methodological background on confidence calibration, not a pallet-pickup test.
- ifm O3R documentation: volCheck. Supports the described volume-check outputs and application examples, without establishing a universal load-support or protective-function claim.
- MiR: MiR1200 Pallet Jack product page. Current manufacturer scope statement for EU-pallet handling and AI-based perception; not evidence of compatibility with every carrier construction.
