2D vs 3D LiDAR for AMRs Which Obstacles Will Your Robot Miss

Should an AMR Use 2D LiDAR, 3D LiDAR, or Both?

The buying decision behind 2D vs 3D LiDAR for AMRs is often presented as a comparison of range, price, and point count. Those specifications cannot answer the most consequential question: will this particular robot detect the objects that can intersect its body or payload, early enough to respond?

Consider an illustrative warehouse route containing a low steel fork, a carton projecting beyond a pallet, and a narrow opening beside a conveyor. A robot may localize accurately along the route while its obstacle representation misses one of those features. Successful mapping and successful collision prevention are separate claims. Procurement teams need evidence for both.

The question remains timely. Hokuyo's industrial robotics comparison was reviewed on September 14, 2026, while its Automate 2026 material highlighted compact multi-layer sensing for mobile equipment. These are signs of continued product and application attention, not proof that every warehouse needs a sensor upgrade. Hokuyo's current comparison provides a useful starting point; this guide develops the purchasing decision around installed coverage and measurable acceptance.

For the wider roles of cameras, encoders, inertial sensors, and scanners, see our mobile robot sensor selection guide. Here, the scope is narrower: deciding whether additional spatial information solves a defined detection problem and proving that the proposed configuration actually uses it.

Write the Obstacle Requirement Before Comparing Sensors

A useful AMR LiDAR selection process begins with the robot's moving envelope. Record the body, top module, carried load, towing arrangement, and permitted movements. The outline changes when a pallet projects beyond the deck, a lift rises, or a trailer swings through a turn. A sensor that protects the empty vehicle's route may leave a loaded corner outside the observed area.

Next, describe targets by geometry rather than warehouse labels. “Detect pallets” is too broad. A pallet presents blocks, openings, boards, wrapping, and a load profile that vary with orientation. “Detect a worker” is a safety-related requirement that needs an appropriate risk assessment and validation; it cannot be reduced to recognizing a convenient test object.

For ordinary object-avoidance performance, define the smallest relevant target dimensions, its height interval, surface characteristics, presentation angle, and required response. Also identify whether it is fixed, appears after mapping, moves into the route, or becomes visible only after an occlusion. These descriptions make competing proposals comparable.

This is a proposed procurement worksheet, not a prescribed safety test. Each row should become a project requirement with an owner, measurable conditions, and a defined result. It also exposes avoidable scope changes: if a supplier priced a flat-floor tote route but the buyer later introduces elevated loads and reversing movements, the original sensing proposal must be reconsidered.

Why a Robot Can Navigate Correctly and Still Miss an Obstacle

A scan plane answers a limited geometric question

A planar scanner observes a slice through the scene. If an object does not intersect that slice, improving the accuracy of measurements within the slice does not reveal it. This is the central issue behind 2D LiDAR limitations, and it should be examined before choosing a longer-range model.

Imagine a horizontal scan plane 180 millimeters above a level floor. An isolated rigid object whose entire height is 80 millimeters does not intersect that ideal plane. A carton projecting into the route only between 600 and 900 millimeters also sits outside it. These dimensions are illustrative geometry, not recommended mounting values or universal detection limits.

Mounting lower can expose some low targets but may create different trade-offs: wheel spray, floor returns, mechanical protection, or occlusion by the robot itself. Tilting a planar sensor changes the heights sampled at different distances. It does not create continuous volumetric coverage. The correct question is which target arrangements the revised installation can demonstrate.

Mapping features and collision targets need different evidence

Walls and rack uprights can provide strong localization features while a narrow projecting object remains poorly observed. Conversely, a robot can detect an obstacle but lack enough stable geometry to localize confidently. Neither result proves the other.

When reviewing a demonstration, ask for the localization display and the obstacle representation separately. Identify which stream influences route planning and which system can command a protective stop. A convincing map should not be accepted as evidence of complete AMR obstacle detection.

For overhanging obstacle detection, the buyer should compare the obstacle's height with every part of the loaded vehicle that can reach it. A horizontal projection into a two-dimensional planning map may conservatively block a route, but height-aware passage underneath requires a more specific model and validation. Simply ignoring everything above the scanner height is not a defensible clearance rule.

A Mounting-Height Calculation That Can Change the Purchase Decision

The following calculation is an original, simplified screening example. It considers unobstructed geometry on a level floor and an ideal downward-looking boundary ray. It ignores beam divergence, target reflectivity, discrete sampling, pitch, vibration, processing, and safety integrity. Its purpose is to identify a question for testing, not to certify detection.

Let the sensor height be H = 0.60 m. Let the lowest boundary of its useful vertical field point downward by alpha = 15 degrees. For a target whose top is at h = 0.10 m, the horizontal distance at which that boundary reaches the target-top height is:

x = (H - h) / tan(alpha) = 0.50 / tan(15 degrees) = 1.87 m

Under these assumptions, the low target lies below the geometric field at shorter distances. Detection farther away may allow a system to track it during approach, but that does not solve the case of an object introduced inside the unobserved region. Prediction and memory must not silently substitute for current observation where the application requires it.

Now place the sensor at H = 0.30 m and give its useful lower boundary a downward angle of 30 degrees. The same calculation becomes x = 0.20 / tan(30 degrees) = 0.35 m. This changes the geometric near-field opportunity substantially, but it does not prove that the second installation detects the target reliably.

The comparison explains why LiDAR mounting height and the distribution of vertical angles can matter more than the maximum advertised range. A high sensor may see beyond a load but overlook nearby floor-level objects. A low sensor may observe those objects while losing sight behind the same load. Mounting changes redistribute visibility.

There is a second reference point to check. Sensor-to-object distance is not necessarily vehicle-to-object clearance. If the leading collision surface is 0.25 meters ahead of the sensor, an object 0.60 meters from the sensor leaves only 0.35 meters of forward clearance in this straight-line example. Turning requires a swept-envelope calculation rather than this simple subtraction.

Ask the supplier to deliver a side-view and plan-view coverage drawing using the actual vehicle datums. The drawing should show useful detection regions, body shadows, load shadows, and the required response boundary. It should distinguish geometric visibility from manufacturer-specified detection performance. A colored field-of-view cone alone does not establish the latter.

Three-Dimensional Data Still Has Gaps

The vertical field may be narrower than buyers expect

The label “3D” does not specify how much vertical space is observed. In its Automate 2026 overview, Hokuyo describes the UCT-10LCM with a 100-degree horizontal and 6-degree vertical field of view. That is a useful example of why purchasers must examine the actual sensing pattern rather than infer broad vertical coverage from the category name.

For 3D LiDAR for mobile robots, request the vertical angle distribution, orientation, selected scan mode, and valid field after installation. A nominal full horizontal rotation can still be blocked by a mast, enclosure, conveyor module, or payload. A different scan mode may change the measurements available where the target appears.

Small targets can fit between useful samples

For a simple illustration near the horizontal plane, beams separated vertically by 2 degrees are about 3 × tan(2 degrees) = 0.105 m apart in height at a horizontal distance of 3 meters. A narrow object can therefore be poorly sampled depending on its position and orientation. Real scanning patterns differ, so this is a screening calculation rather than a product performance claim.

Point count thresholds can widen that gap between a visible return and an actionable detection. Ouster's Zone Monitor API documentation explains that a zone cannot trigger when its required point count exceeds the number of beams intersecting the zone. It also explains that an occluded region cannot be reliably evaluated for occupancy. These are concrete examples of why coverage and trigger logic must be reviewed together.

Minimum range is specific to the configuration

Evaluate 3D LiDAR near-field detection for the delivered configuration; do not assume every 3D sensor has the same limitation. Ouster's minimum-range documentation distinguishes hardware generations and configurable thresholds. The purchasing task is to verify the delivered hardware, firmware, reporting threshold, mounting obstruction, and downstream filtering.

Describe AMR blind spots as separate mechanisms: outside the optical field, blocked by an object, too sparsely sampled, suppressed before transmission, discarded by software, or observed too late for the required response. Different mechanisms need different remedies. Replacing a sensor cannot reliably fix a target that a software filter intentionally removes.

Follow the Obstacle Through the Software, Not Just the Point Cloud

A supplier can show a bright cluster in a visualization while the planner receives an empty cell. The measurement may be filtered by height, removed as floor, rejected as noise, assigned an old timestamp, or transformed into the wrong coordinate frame. Acceptance must follow the information from acquisition through the commanded response.

For example, the official Nav2 Voxel Layer documentation describes a three-dimensional internal representation projected into a two-dimensional planning costmap. It exposes height, marking, clearing, and occupancy controls. This illustrates how software settings determine which observed structures influence navigation. It does not establish that any particular Nav2 configuration is safety-certified.

During an acceptance review, select one difficult target and inspect four synchronized records: the original measurement, the filtered representation, the planning obstacle, and the vehicle response. Repeat after changing the target's height or approach angle. If it disappears between stages, investigate that transition before increasing hardware cost.

Time must remain visible throughout this chain. At an illustrative travel speed of 1.0 meter per second, an additional 0.15 seconds of sensing and processing delay corresponds to 0.15 meters of travel before the resulting decision. This is not a stopping-distance calculation; braking, control response, uncertainty, and other applicable allowances remain separate.

Require latency measurements under representative computing load, including logging, camera processing, map updates, and communications. An average from an idle robot hides occasional long delays. Preserve the distribution and the conditions producing the largest observed values. The permitted response time belongs in the requirement before the demonstration begins.

Compare Complete Architectures, Including Targeted Additions

A buyer usually has more than two choices. Retaining an effective localization scanner while adding a sensor for a documented gap can be more practical than replacing the navigation stack. Alternatively, a new platform with integrated volumetric perception may be preferable when the operating environment changes frequently.

The choice between LiDAR and a depth camera for an AMR should use the same target requirements. A depth camera may offer useful close-range geometry or appearance information; its suitability depends on the technology, illumination, target surface, mounting, and processing. A LiDAR proposal likewise needs testing against the actual target and environment. Neither label is a universal answer.

For low obstacle detection, compare the useful coverage immediately around the vehicle, not just a picture of a distant object. Test arrival from the side and emergence from behind a load. If a supplementary device is meant to close a specific gap, require the supplier to identify that gap on the coverage drawing and demonstrate the relevant transitions.

Environmental performance is a separate test axis. Glass, dark materials, glare, contamination, and condensation may expose weaknesses that clean geometric targets do not. Use the scenarios in our AMR sensor environmental testing guide to extend the comparison without assuming that more dimensions remove material-dependent limitations.

Keep Personnel Protection Explicit in the Purchase Specification

The dimensionality of a sensor does not establish the integrity of a safety function. A navigation point cloud, a collision warning, and a validated personnel-protection function require different evidence. The contract should state which functions are credited for risk reduction and which improve operating performance.

Three-dimensional safety devices do exist. SICK's safeVisionary2 product overview, for example, describes a safety-rated 3D time-of-flight camera. That product category must not be conflated with an ordinary depth camera or a generic 3D LiDAR. Its permitted use still depends on the exact function, configuration, operating limits, and complete application.

Conversely, Ouster's Zone Monitor documentation explicitly limits safety-critical use for the firmware and earlier sensor generations identified there. A feature capable of reporting zone occupancy is not automatically qualified to protect people. Verify the current safety manual and certification scope for the exact equipment offered.

Our guide to safety-rated sensing versus navigation perception explains that evidence boundary in more detail. For this purchase, require a clear allocation of responsibilities between the perception layer, protective system, motion controller, and integrator. Do not reduce a protective field or disable a protective function simply to make a demonstration pass.

Run an Obstacle Challenge That Competing Suppliers Can Reproduce

The following test framework is an original proposal for comparing object-detection performance. It complements an application-specific safety validation program; it does not replace it. Conduct physical trials in a controlled test area with appropriate supervision and representative test objects, keeping people out of experimental collision paths.

Freeze the test baseline

Record the robot, sensor models, firmware, mounting coordinates, perception configuration, software versions, payload, tire condition, permitted speed, and map. Record the actual target dimensions and surface properties. Photographs should show both target placement and sensor installation so another team can reconstruct the setup.

Any adjustment during testing creates a new configuration. Retain the failed result, document the adjustment, and repeat the affected conditions. A supplier should not quietly change a filter between demonstrations and combine the best results into a single acceptance claim.

Challenge the geometry in more than one direction

Use enough repeats and condition combinations to support the particular claim. A short demonstration can expose a coverage gap; it cannot prove an extremely low failure probability. Choose the sample plan before seeing the results, and preserve unsuccessful trials as well as successful ones.

Measure what happens before the collision boundary

Useful records include the earliest sustained target detection, target identity, measurement timestamp, first planning-layer representation, commanded response, actual vehicle motion, and closest clearance. State the reference frame for every distance. Add whether an independent protective function intervened, since its intervention can hide a failure of the navigation perception being evaluated.

Define what “sustained” means for the project. A single return that disappears may not support the intended response; requiring too many consecutive observations can delay it. The acceptance rule must account for both detection continuity and elapsed time. Keep the chosen rule consistent between suppliers.

Record false obstacle events too. An architecture that avoids every test object by treating most of the warehouse as blocked may satisfy a narrow demonstration while failing the production task. Track manual interventions, unwanted stops, route completion, and delays under the same conditions. Productivity evidence belongs alongside the detection evidence.

Connect the resulting records to the wider AMR site acceptance test matrix. Assign each deviation an owner and an approved disposition. Acceptance should describe a released configuration and operating envelope, rather than a general statement that the supplier's sensors worked.

Price the Complete Detection Capability

Sensor unit price is only one part of the commercial comparison. Include mounting changes, protective housings, compute capacity, power and networking, software integration, calibration, validation, spares, cleaning access, and future revalidation. Ask which items are included in the quotation and which become engineering change orders.

Compare configurations that meet the same requirements. A low-price proposal that excludes reverse travel, certain loads, or low obstacles is not equivalent to a more complete offer. Those restrictions may still be commercially acceptable, but they must be visible and operationally enforceable.

For an illustrative operating-cost model, suppose records attribute eight interruption events per working day to the specific gap under review. If each consumes four minutes of robot unavailability, that is 32 robot-minutes per day, or 133.3 robot-hours across 250 working days. This arithmetic describes robot unavailability, not automatically lost factory output.

If a properly controlled trial subsequently demonstrates a 50 percent reduction in those attributable minutes, the modeled improvement is about 66.7 robot-hours annually. That percentage is an assumption for illustration, not an expected sensor benefit. Convert time to money only after considering spare fleet capacity, labor involvement, process bottlenecks, and whether demand was actually unmet.

Request a sensitivity analysis for the recurring problem's frequency and the intervention cost. Include maintenance or computing burdens introduced by the proposed upgrade. The business case is strongest when the supplier can connect additional sensing to a measured problem, a demonstrated improvement, and a maintainable configuration.

Finally, establish who owns future changes. A replacement bracket, taller payload, revised filter, or different firmware can alter the accepted coverage. Use AMR change control and revalidation to decide which evidence must be renewed before the configuration returns to production.

Make the Decision Conditional on the Evidence

Retain an existing 2D-centered architecture when the relevant geometry is adequately observed, permitted operating restrictions are practical, and testing demonstrates the required performance. An upgrade has little value if it adds data without improving an application requirement.

Add targeted or broader 3D perception when a documented height-dependent gap remains and the proposed installation closes it under representative conditions. Confirm that software preserves the additional information and that response timing remains acceptable. Keep the protective system's responsibilities independently defined.

Redesign the application when sensing alone cannot provide a dependable operating envelope. Load presentation, route layout, viewing position, vehicle geometry, or the material-handling process may need to change. A procurement decision can legitimately end with a different application design rather than a more expensive sensor.

The final purchase record should identify the targets covered, the conditions excluded, the evidence reviewed, the remaining restrictions, and the owner of each unresolved item. That record gives operations something more durable than a specification sheet: a clear account of what the delivered robot has been shown to detect.