One inventory cart reports a fallen box in aisle 14. Twenty seconds later, a second cart approaches from the opposite end and reports an obstruction near the aisle entrance. The warehouse server now has two plausible descriptions and an awkward question: did the carts see the same box?

If it assumes two obstacles, it may close more of the aisle than necessary. If it merges two separate obstructions, it may assign work beyond a blockage it has overlooked. Both descriptions can be accurate while leaving the coordinator unable to make a useful distinction.

This fictional fleet counts fully visible cartons on single-depth shelving. Labels identify the shelf and expected SKU; hidden stock remains uncounted. Each cart performs its own observations, while an on-site server coordinates assignments. The server’s decisions depend on the information those observations preserve as they cross from perception into the application.

InformationAmbiguous handoffHandoff designed for comparison
Position and extent“Near the entrance”Location and occupied area in a named map frame, in stated units
TimeMessage arrival timeCapture time, with the clock reference and its known limitations
UncertaintyOne confidence scoreSeparate information about classification, geometry, and observation quality
Evidence

A sentence without a sourceSensor and observation identifiers, with a retrievable evidence reference where permitted

Preserve the distinctions the next decision needs

The sentence “a fallen box is blocking aisle 14” is useful to a person glancing at an alert. For a coordinator comparing reports, it leaves substantial work unresolved. Where along the aisle is the box? How much space does it occupy? Did the second cart see it recently, or did its report spend time waiting for a connection?

An accompanying image can help answer some of those questions, provided the recipient can interpret its geometry and capture conditions. A language model may also extract a richer description. Neither approach automatically supplies a shared position or establishes that two observations concern the same object. Those relationships need to be represented and evaluated explicitly.

The handoff therefore needs an agreed specification for the information supplied and the behavior expected when it is incomplete. Software engineers often call this an interface contract. Its value is practical: the receiving application should be able to distinguish a valid position from an unknown one without guessing what the sender meant.

For the warehouse, the report needs to preserve the object’s location and extent in a named coordinate frame, with stated units. It also needs the capture time, separate indications of uncertainty, and a way to trace the report back to its source observation. A readable summary can sit alongside that information for the supervisor.

Put every position in a frame the receiver understands

A detector commonly locates the box within a camera image. Those pixel coordinates describe where the box appears on the image plane. The coordinator needs a location on the warehouse map, and the conversion requires more than knowing where the cart is standing.

The quality of decisions that are made using camera inputs depends on how we configure and process the data produced. A range measurement, or a valid geometric constraint such as a calibrated floor plane, is then needed to locate the relevant points in three-dimensional space. The cart’s estimated position and orientation at capture time allow those points to be transformed into the map frame. A box’s extent needs suitable geometry as well; the apparent width of its image alone does not establish the space it occupies on the floor.

Every part of that conversion can carry error. A depth reading may be unavailable, the floor-plane assumption may fail, or the cart’s location estimate may have drifted. The application should preserve those limitations. If the available observation supports only a region where an obstacle might be, the report should describe that region rather than present a precise point without justification.

The ROS project’s REP 103 documents standard units and coordinate conventions precisely because inconsistent conventions cause integration errors. A camera’s optical frame and a robot’s body frame, for example, use different axis orientations. Naming the frame and documenting the transformation prevents two components from interpreting the same numbers as different physical locations.

Once both carts report in a shared map frame, the server can compare the estimated occupied regions. Overlapping locations and compatible dimensions provide evidence for a common obstacle. They support an association hypothesis; they do not prove identity. Two similar boxes could be close together, or a worker could move one between observations. The coordinator should retain unresolved alternatives when the evidence does not justify merging them.

Keep observation age and limitations visible

Capture time tells the receiver when the scene was observed. Processing completion and message arrival describe later events in the pipeline. They remain useful for diagnosing delays, but using arrival time as the observation time can make an old scene appear current.

Comparing capture times across carts requires clocks with a known relationship. If clock error is large relative to the event being analyzed, apparent timing differences may be misleading. The contract should make the time reference clear and define when an observation has become too old for a particular decision. An obstacle report may remain useful for scheduling after it has become unsuitable for deciding whether a narrow route is currently clear.

Uncertainty also needs more than a single score. The system may be confident that an object is a carton while having a poor estimate of its distance. A precise depth reading may accompany an uncertain object label. Glare may affect part of an image without invalidating every measurement from it. Combining all of these conditions into one confidence value leaves the receiving application unable to respond to their different consequences.

For inventory, the distinction is especially tangible. A shelf label can establish which SKU belongs at a location; it cannot reveal cartons hidden behind an obstruction. The count should cover what the cart could actually observe. An unobserved portion of the shelf remains unknown, and another observation or a manual count may be needed to complete the task.

Provenance makes such questions possible to investigate. The report can identify the cart, sensor, and source observation, along with the processing version used to produce it. Where access and retention rules permit, a reference can lead to an image crop or recorded frame. That evidence need not travel with every message, but its availability and eventual expiry should be understood by the recipient.

Choose the fusion boundary around the measurements

Additional sensors introduce more measurements to reconcile. A camera may suggest the object’s type while radar contributes range and radial velocity. Those readings can be combined only with appropriate spatial calibration, temporal alignment, and an assessment of whether they concern the same target. A radar return with no reliable camera match can remain an object of unknown type; inventing a label would conceal the disagreement.

Fusion can take place at different stages, depending on what the downstream task needs and what each processor can handle. Sometimes individual sensors can provide compact detections. Other applications benefit from combining richer measurements before reducing them. The choice should account for what information would become unavailable after that reduction, together with the cost of moving and processing it.

A concrete example appears in Jason Huang’s account of Ambarella’s radar demonstration. Huang describes a CV3 system-on-chip in the front radar unit processing data from five Oculii radar heads, then sending a combined point cloud to a zone controller for sensor fusion. A point cloud represents individual detections by their positions in three-dimensional space. That design keeps multi-radar processing together and gives the next stage a spatial representation. Other fusion tasks may need a different boundary, depending on which measurements the receiving stage requires.

Test what the interface allows the receiver to conclude

The warehouse contract can be tested by replaying reports from known scenes. Start with two observations of one box, then use two nearby boxes with similar appearances. Introduce an old report after a newer one. Remove the range estimate, change the coordinate frame, and simulate a clock offset. The expected result should include how the coordinator handles uncertainty in each case.

A useful test checks the downstream decision. If two reports cannot be associated confidently, the coordinator should preserve that ambiguity and follow the specified scheduling behavior, perhaps requesting another observation. A missing distance must not become zero, and a region hidden from the camera must not become clear space. Passing a message through a parser establishes much less than this.

Run these checks on the target hardware with the capture and processing workloads active. Confirm that timestamps refer to the intended exposure and that transformations use the cart’s pose at that time. Check whether queueing changes the age of information presented to the receiver. The interface may be logically consistent while still delivering information too late to use.

The coordinator’s assignment remains a proposal evaluated by the cart’s application runtime; motion remains subject to the machine’s separately validated protective control. A well-designed report improves coordination without taking over those responsibilities. When the two carts report from opposite ends of aisle 14, the server should have enough evidence to make a justified association—or enough information to recognize that another observation is needed.