How to Address the AMR Perception Gap

Warehouses and manufacturing facilities expose fundamental challenges for conventional AMR perception systems. Unlike the simple layouts and ideal conditions of controlled pilot projects, most facilities are cluttered, dynamic and visually inconsistent, with frequent occlusions, dust and fluctuating lighting. In these circumstances, perception failures tend to concentrate in the most critical region: the near-field zone, where robots navigate around objects, infrastructure and people.

Many systems rely on 2D and 3D lidar and various types of 3D cameras, but each modality presents trade-offs in performance, reliability or cost.

Most lidar systems use a method called “time of flight” (ToF) in which the sensor measures distance by calculating how long it takes a laser pulse to travel to an object and back. The lidar then uses the data from the returned signals to create a “point cloud,” which in 3D looks topographic and in 2D appears as dots on a plane.

Whether 2D or 3D, lidar has trade-offs for AMRs. 2D lidar scans in a single horizontal plane, producing a thin slice of distance measurements suited for basic navigation and obstacle avoidance on flat ground. It can miss overhanging objects or open loading bays because it does not capture full 3D structure. 2D lidar can also struggle in warehouse environments where there are repetitive patterns and similar-looking aisles.

3D lidar captures depth across multiple vertical angles to generate a full 3D point cloud, enabling much richer understanding of shapes, height and complex scenes. While 3D lidar can solve many perception challenges and often is used for ground truth and training, its higher cost and system complexity — particularly for mechanically scanning systems with moving parts that can wear out over time — can limit scalability for AMRs, although newer solid-state designs are reducing these constraints.

Camera Systems Show Part of the Picture

Camera systems provide high-resolution visual detail and are essential for recognizing objects and features. However, they do not inherently measure depth. Stereo-vision, structured-light, and ToF cameras can estimate depth, but their performance degrades in low light, glare, low-texture environments or when surfaces are reflective or absorptive. All vision-based systems are also constrained by narrower field of view and resulting blind spots.

While 3D camera solutions are generally less expensive than lidar, they still introduce meaningful cost and integration complexity. A single RGB camera is far more cost-effective but requires AI or other sensor fusion to infer depth, resulting in additional software and the powerful compute required to run it.

Both lidar and camera systems also struggle with detecting thin or protruding objects — the so-called “pokey problem” — such as pallet edges or forklift tines. Dust, transparency and reflective surfaces can further degrade lidar performance as well, often resulting in false positives that trigger unnecessary stops and reduce operational efficiency.

n additional issue is that lidars can be confounded in a warehouse, where there are often pallets shrink-wrapped in plastic cling film or shiny epoxy floor coatings. The glossy, transparent and reflective nature of these surfaces scatters, distorts and reflects laser beams, producing false distance measurements and broken perimeter models.

Radar Is a Key Component

Radar overcomes many of these limitations by employing radio-frequency electromagnetic waves, which have longer wavelengths than the visible or infrared light used by lidars and cameras. As a result, it is inherently insensitive to lighting conditions and far more robust to environmental variability. Radar can also detect objects with low visual contrast and, in some cases, objects that are partially obscured or encased in reflective materials that can confuse lidar.

Modern radar technology has matured significantly, driven by automotive adoption. Newer 4D imaging radar systems measure range, azimuth, elevation and velocity simultaneously, providing both detection and motion awareness. Radar that is specifically designed for high-precision 360° detection around a vehicle, with a wide field of view and a range of up to 10 meters — considered ultrashort by automotive standards — is particularly well suited for near-field warehouse applications. Because radar is produced at automotive scale, it also offers a compelling cost advantage.

Putting It Together

The Aptiv PULSE™ Sensor combines the strengths of radar and vision into a compact, cost-efficient perception system. It integrates a single wide-angle RGB camera for visual context with radar that can achieve precise, near-field sensing for depth and motion sensing. The radar provides reliable detection and tracking in challenging conditions, while the camera contributes semantic understanding and fine spatial cues.

The system is compact — roughly the size of an ice cube — enabling reduced material costs and fewer constraints on design choices.

PULSE is optimized for tasks that demand both precision and reliability, such as detecting small protruding objects or navigating tight, cluttered spaces. In these scenarios, its radar ensures that objects are consistently detected, and its camera preserves the benefits of optical detection. By combining complementary sensing modalities, PULSE provides a robust and scalable perception solution that enables true autonomy — allowing robots to navigate independently, operate more safely around people and perform tasks without constant human oversight.