Vision target routing across overlapping robot work areas

A camera can see a part that a particular robot cannot pick. If distribution considers workload before reachability, valid detections can become rejected robot targets. Cropping the image may hide the symptom while leaving part of the conveyor uninspected.

For conveyor-cell designers and commissioning teams, the better approach is to separate image coverage, target eligibility and queue ownership. Keep the required belt coverage, transform detections into calibrated coordinates, and route each physical part only to eligible robots. This explainer develops that logic, including overlap regions, duplicate suppression, unreachable targets and the separate controls required for collision protection.

Keep image coverage separate from robot eligibility

The camera's detection region answers one question: where should the system look for parts? A robot's pick region answers another: which detected poses can that robot service under the current operating conditions?

Those regions need not have identical boundaries. One full-width image can supply a common target stream, while downstream classification assigns different targets to different robot queues.

Start with two distinct stages:

• Eligibility: determine which robots can service the target within its available pick window.
• Allocation: choose an owner from that eligible set using workload, sequence or another production rule.

Load balancing belongs after eligibility. An idle robot is not a valid destination merely because its queue is shorter.

This differs from configuring which request selects which target source. The guide to target-selection mapping and rejected requests addresses that interface. Here, the central question is whether the selected robot can actually service a particular physical part.

Define pick regions in calibrated coordinates

Classify targets in a shared conveyor coordinate system rather than splitting the image at an arbitrary pixel column. Perspective, camera rotation and lens distortion can make an image-space boundary differ from the intended physical boundary.

The camera calibration guide for length measurement explains the underlying image-to-world mapping. Routing also requires consistent transforms between that reference frame and each robot's operating frame.

Define each eligible region using validated pick poses, not just nominal arm reach. Its limits can depend on:

• Part orientation, height and grasp offset.
• Tool dimensions and the required approach and departure paths.
• Robot configuration and applicable joint limits.
• Conveyor position at the planned interception time.
• Calibration uncertainty and target localization error.

A rectangular cross-belt range can be sufficient for a simple installation. Irregular workspaces may need polygons, pose-dependent lookup tables or the controller's documented feasibility checks. A position inside a polygon is only a candidate if the approach and subsequent motion remain feasible.

Do not invent configuration parameters to implement this abstraction. Use the documented routing, region-filtering and queue facilities available in the installed system.

Predict where the part will be when picked

A detection describes a part at acquisition time. Routing must account for its position when the robot can intercept it, especially when robots occupy different positions along the conveyor.

For a straight conveyor, define x along belt travel and y across the belt. If encoder displacement is expressed in the same calibrated distance units, the basic estimate is:

Here, s(t) is conveyor displacement. The estimate assumes the part follows the belt without lateral movement or slip. Orientation and height must also be tracked where they affect the grasp.

For example, a target detected at x = 100 mm moves another 250 mm before interception:

If a robot's validated longitudinal pick interval is 300–800 mm, that predicted position passes the longitudinal check. It must still pass cross-belt, orientation, timing and motion-feasibility checks.

At a constant 250 mm/s, the same displacement takes 1 s. During acceleration, use encoder-based tracking rather than assuming that speed remains constant. The guidance on sharing a conveyor encoder covers the common motion reference, while scene timing and stale pick data addresses target age.

Work through an overlapping-region example

Consider an illustrative conveyor with a width of 1,000 mm. Robot A has a nominal cross-belt pick range from 0–600 mm, and robot B from 400–1,000 mm.

Suppose commissioning establishes an inward routing allowance of 20 mm at each boundary. The resulting candidate ranges are:

The allowance is an example, not a universal recommendation. Derive yours from the error budget and validated operating envelope, including tool and motion constraints.

For targets eligible for both robots, choose a nominal ownership boundary at y = 500 mm. Assign overlapping targets below that boundary to A and targets at or above it to B. Apply this tie-break only after eligibility testing.

These results assume all other eligibility checks pass. If B cannot meet the target's pick window, the target at 500 mm may be reassigned to A through a controlled ownership transfer.

Full-width imaging does not establish full-width picking capability. In this example, the outer strips remain visible but fall outside the candidate ranges. Recovering those parts requires a validated change in presentation or coverage, not a smaller camera region that hides them.

Choose where routing should happen

There are several reasonable architectures. The choice depends on the documented capabilities of the vision and robot-control software, especially its queue and rejection behavior.

A central coordinator can detect each physical part once, calculate its eligible set and issue one assignment. This makes shared-region ownership explicit, but it requires a defined recovery process if the coordinator restarts.

Spatial classification can also work. Assign a routing class from calibrated position, then map that class to an appropriate destination. If separate detection models have overlapping search regions, reconcile their results into one physical track before assignment.

Robot-side filtering is appropriate only when the documented queue behavior supports it. A rejected target might be deleted, returned, deferred or passed downstream; these outcomes are not interchangeable. Verify the actual behavior before relying on rejection as a handoff mechanism.

Give each physical part one owner

Spatial eligibility alone does not prevent two robots from receiving the same part. Duplicate detections and duplicate assignments are separate problems.

Duplicate detection occurs when overlapping models or successive images report the same physical object more than once. Duplicate assignment occurs when one valid physical track is claimed by multiple consumers.

Use a persistent track identity where the platform supports one. A per-image detection number is not automatically a physical-part identity across images. Associate repeated observations using conveyor displacement, predicted position, object appearance and other validated tracking information.

A useful ownership lifecycle is:

The exact states depend on the platform. The essential requirement is that a claim has one authoritative outcome. If multiple consumers compete, reservation must be atomic or provide an equivalent guarantee.

For example, if two robots request the same track within 10 ms, only one reservation should succeed. A cancellation or ownership transfer must be acknowledged before another robot receives permission to act. Delayed messages from an old assignment must not revive it.

Keep terminal track records long enough to reject delayed duplicates over the maximum relevant processing and communication delay. Validate that retention interval from measured system behavior rather than choosing it solely from frame rate.

Handle unreachable, busy and expired targets differently

An empty eligible set means no robot currently meets the target's pick conditions. That is different from having eligible robots that are temporarily occupied.

Record explicit outcomes:

• Outside validated coverage: no robot can service the pose under the applicable conditions.
• Eligible but busy: a robot could service it, but cannot currently accept the assignment.
• Expired: the remaining pick window is too short for a valid interception.
• Invalid target: pose, identity, calibration or tracking quality is insufficient.

Give each outcome a deliberate production response. Depending on the cell design, that may mean a downstream station, a controlled reject path, an alarm or an approved conveyor-control action. Do not silently discard the target or circulate it indefinitely between queues.

A target visible upstream may still reach a downstream robot later. Conversely, a target inside a geometric region may already be too late to pick. The explanation of conveyor target availability and cycle-time losses helps distinguish these timing cases.

For targets waiting on capacity, define a latest assignment time and evaluate it continuously. If conveyor speed changes are part of the approved process, coordinate them through the documented controls described in vision queue conveyor speed control.

Select imaging and processing hardware independently

Routing is primarily a coordinate, tracking and queue-management problem. A different camera does not by itself correct an ownership error or an incorrect robot transform.

Choose area scan cameras for the required belt coverage, image quality and acquisition interface. Use the camera finder to compare imaging options after establishing the smallest feature that must be located reliably.

The LS-CMR-0395 is a monochrome GigE area scan camera. Whether it suits the inspection depends on the application's imaging and timing requirements.

Consider vision controllers where dedicated camera connections and centralized processing fit the architecture. Verify software licensing, interface support, tracking functions and queue integration independently of the hardware specifications.

The MV-SVC2506-16G60-128G provides camera GigE ports and built-in light outputs. Those features do not establish support for a particular allocation algorithm or robot interface.

Neither hardware choice establishes compatibility with the other components. Confirm the complete acquisition and control path before selecting parts.

Keep collision protection separate from routing

Two robots may each be able to reach a target while being unable to execute their respective motions simultaneously. Reachability answers whether an individual operation is feasible; collision management considers interacting motions and occupied space.

Maintain the cell's risk-assessed safeguards and documented protective functions. Where shared process space requires coordination, use the controller-supported interlocks and motion coordination appropriate to the validated design.

An overlap-ownership rule can reduce conflicting requests, but it does not constrain every tool, arm or carried-part position. Approach paths, departure paths, recovery moves and stopping behavior all matter.

Likewise, do not suppress controller reachability checks to make rejected targets run. Treat rejection as evidence that the model, timing or assignment needs investigation. When overlap coordination is unavailable or inadequate, use a validated non-overlapping operating arrangement rather than assuming the filtering logic provides protection.

Verify the routing contract before production

Test routing without motion first, using recorded detections or an approved observation mode. Compare the calculated eligible set, selected owner and expected outcome for every test target.

Then verify the complete cell under its approved commissioning procedure. Include these checks:

• Shared coordinates: confirm axis directions, units, conveyor reference and robot transforms.
• Boundary behavior: test targets on both sides of each limit and exactly on the ownership boundary.
• Pose dependence: include the orientations, heights and grasp offsets allowed by the recipe.
• Overlap ownership: confirm one reservation and one final disposition per physical track.
• Capacity changes: make one eligible robot unavailable and verify reassignment or the defined exception path.
• Delayed messages: test stale assignments, duplicate observations and expired targets without allowing unintended motion.
• Restart behavior: define how old reservations are reconciled after communication loss or controller restart.
• Coverage gaps: log visible targets outside the union of validated pick regions instead of masking them.

The guide to conveyor vision startup synchronization covers related restart and initialization checks. Record a trace containing acquisition time, conveyor reference, physical track identity, eligibility decision, ownership changes and final disposition.

Before ordering, document the belt coverage, calibration method, target rate, interfaces and ownership contract. Start the imaging selection in the camera configurator, or contact the applications team with those requirements and the routing trace.