Vision-Guided SCARA Robot: Stabilize Infeed First
Direct answer: a vision-guided SCARA robot is dependable only inside a declared presentation and calibration window. Control how parts enter the camera field, separate image calibration from camera-to-robot calibration, reject coordinates outside the approved pick region, prove that the gripper actually acquired the selected part, and give every uncertain image or transfer a bounded route. Recognition confidence alone is not a production acceptance rule.
The practical design object is an infeed-to-exception map. It begins before the camera triggers and ends after the destination confirms placement or the part reaches a defined reject location. This approach prevents a clear demonstration image from hiding unstable feeding, coordinate drift, ambiguous pickup, or a retry loop that consumes the real takt time.
Define the population for a vision-guided SCARA robot
List the part families and the visual differences that matter to the task. Record outline, height, surface finish, color, reflectivity, permitted orientation, permissible overlap, spacing, contamination, and any features the algorithm must find. If two variants share an outline but require different destinations, define the feature or upstream identity that distinguishes them. A classifier label that cannot be tied to the physical part should not authorize motion.
Presentation is a mechanical requirement. Describe whether parts arrive in pockets, on a conveyor, from a bowl feeder, or in a random layer. Establish the working field of view, depth range, settle behavior, conveyor tracking condition, and border exclusion zone. A feature found at the edge of an image can still yield an unusable grip because the tool body, cable, or part sweep extends outside the robot-safe pick region.
Use samples that represent production surfaces. Bright, dark, scratched, oily, translucent, patterned, or worn parts may respond differently to the same light. The purpose of a challenge set is not to make the algorithm fail; it is to locate the boundary where the result must change from “pick” to “hold, retry, or reject.”
Keep the calibration chain visible
Cognex’s guided pick-or-place documentation describes image acquisition, feature detection, alignment offsets, robot correction, image calibration, and hand-eye calibration as connected but distinct functions. Preserve that distinction in the project record. Image calibration relates pixels to a meaningful image-space measurement. Hand-eye calibration relates the camera frame to the robot or work frame. A good image does not prove that the transformed robot point remains valid.
Write the coordinate chain from detected feature to commanded pick: image frame, calibrated image plane, camera frame where applicable, conveyor or fixture frame, robot base, tool frame, and the active grip reference. Record which transform changes after a camera move, lens adjustment, robot mastering event, gripper replacement, fixture service, or software restore. Define a quick verification artifact and a tolerance for deciding that production may continue.
Part height deserves a separate line. A two-dimensional location can be correct in X, Y, and rotation while the top surface or pickup plane changes enough to compromise clearance or grip. If height is constrained mechanically, prove that constraint. If it varies, specify the sensing or compliant method used to keep the approach safe and the pickup reliable.
Turn a detection into a permitted pick
The vision output should carry more than X, Y, and angle. Define a result record containing part class, feature score, pose, timestamp or trigger identity, calibration revision, image or inspection reference where retained, and any limit flags. Before movement, controls should confirm that the result is current, inside the permitted pick region, associated with an available physical part, and compatible with the active recipe and gripper.
A useful acceptance table is specific to the application:
| Result condition | Robot action | Part disposition |
|---|---|---|
| valid class, pose, calibration, and pick region | execute the approved approach and pickup | continue to grip proof |
| low confidence but image remains usable | apply a bounded reacquire rule | wait without advancing identity |
| pose outside motion or tool limits | inhibit pick | route or request feed correction |
| overlap prevents a unique target | do not guess between parts | separate, recirculate, or reject |
| calibration check expired or failed | block transformed moves | hold the station for verification |
Avoid endless reacquisition. Set a maximum number or time budget, then move to a known disposition. Otherwise a rare image problem becomes an unmeasured queue that distorts output and leaves operators improvising at the station.
Prove grip and placement independently
The robot reaching the pickup point does not show that the part left the infeed. Select evidence appropriate to the end effector and part: vacuum pressure behavior, jaw position, part-present sensing, weight or force response, downstream nest confirmation, or a second visual check. Understand what each signal can and cannot prove. Vacuum present might indicate a blocked cup rather than the correct part; jaw closed might mean no part was acquired.
Define the transfer envelope for the held part, not only for the empty tool. Include possible part rotation, flexible or protruding features, and the consequence of a dropped or double-picked item. If grip evidence disappears in flight, specify a controlled stop or disposal path that avoids releasing the part over equipment or people.
Placement needs its own completion rule. The command can finish while a part remains on the tool, rebounds from a pocket, or occupies the wrong destination. Decide whether the station needs release confirmation, destination occupancy, orientation verification, or downstream acceptance. Preserve the part identity through the handoff so a reject result cannot attach to the following cycle.
Design exceptions before optimizing nominal motion
Map at least five non-nominal routes: no part found, multiple candidates, result outside limits, pickup not proved, and destination unavailable. Add stale result, calibration invalid, feeder jam, camera unavailable, and human intervention where relevant. Each route should name the state owner, time limit, allowed retry, physical disposition, operator information, and restart prerequisites.
The interface can be drawn as a series of custody changes. The feeder owns presentation until the inspection window is stable. The vision task owns a result only while its trigger and calibration are current. The robot owns transfer after pickup is proved. The destination owns the part after placement acceptance. This prevents two controllers from simultaneously treating the same location as free or the same part as complete.
NIST frames robot performance assessment around observable requirements, metrics, and repeatable test methods. Apply that principle by measuring the full decision chain: presentation success, valid-image rate, false acceptance, missed target, coordinate residual, pickup proof, placement acceptance, retries, rejects, and recovery time. Do not report only algorithm accuracy if feeding or grip losses dominate production.
Build a vision challenge matrix
Create trials across the variables most likely to interact. Use the lowest and highest allowed spacing, rotation, height, reflectivity, and illumination; include border positions, partial occlusion, neighboring parts, damaged features, and permitted background variation. Repeat after warm-up and after routine service that can affect the camera or gripper.
For each sample, retain the expected class and pose, actual result, transformed target, whether motion was permitted, grip evidence, final destination, and exception code. Review false positives separately from missed detections because their risks differ. A missed part may reduce output; a plausible but wrong target can direct the tool to a collision or send the wrong part downstream.
Challenge the calibration deliberately. Move a verification artifact to known positions, check residual error across the field rather than only at the center, and repeat after a controlled camera or tool service. Include a failed check to prove that transformed motion is inhibited. The acceptance tolerance must follow the real gripper and placement clearance, not an isolated camera specification.
Calculate takt from the event distribution
A cycle budget should distinguish feed, settle, trigger, exposure, processing, communication, robot approach, pickup, proof, transfer, release, destination confirmation, and buffer update. Add the observed frequency and duration of reacquisition, feed correction, reject handling, replenishment, and calibration checks. The average of clean cycles cannot describe a station where occasional overlap creates long recovery events.
Use a queue model when the camera, conveyor, and robot operate asynchronously. Confirm how results are ordered, when a target expires, what prevents double use of an image, and how the system behaves if the robot skips a candidate. If conveyor tracking is used, validate timing and encoder relationships across the approved speed range and after a controlled stop.
Protect calibration, replenishment, and fault-clearing tasks
ISO 10218-2 places robot-cell integration and lifecycle tasks inside the application boundary. ISO 12100 requires systematic hazard identification and risk reduction. For a SCARA sorting cell, examine robot and feeder motion, pinch and crush points, part ejection or drop, sharp parts, stored pneumatic energy, and access during replenishment, lens cleaning, focus or light adjustment, calibration, reject-bin service, jam clearing, and restart.
The protective design should reflect who performs each task and what energy must remain available. A calibration routine may need camera and limited robot functions while automatic feeding is inhibited. A reject-bin change may require the robot to stay away from an access zone. Define validated modes and indicators rather than relying on a general stop button as the complete task procedure.
Application-review inputs
- representative parts covering permitted pose, height, surface, color, contamination, spacing, overlap, and damaged-feature conditions
- feeder or conveyor drawings, speed and settle data, camera and lens, lighting geometry, field of view, exposure limits, and ambient-light conditions
- coordinate-frame definition, calibration procedure, verification artifact, allowable residual, robot and gripper data, pickup and placement clearances
- signal list, result identity, timeout and retry rules, destination capacity, reject route, traceability requirements, and target takt distribution
- operator tasks, replenishment and cleaning methods, safeguarded spaces, fault scenarios, restart rules, maintenance intervals, and governing safety requirements
How this guide was prepared
In practice, the editorial review checked the retained SCARA sequence at 00:02.4, 00:12.0, and 00:24.0. The first view shows a SCARA tool above a tray of colored geometric parts; the second shows the arm displaced from the tray while the presentation area remains visible; the third shows the work area together with a monitor displaying colored shapes. These observations support the article’s focus on presentation, coordinate transfer, pick evidence, and destination handling. They do not establish recognition percentage, pick accuracy, takt, calibration quality, supplier identity, or production acceptance. Official vision, robot-cell, machinery-risk, and measurement sources set the engineering boundary.
Citation-ready statements
- According to Cognex guided pick-or-place documentation, a usable vision result belongs inside a calibrated robot pick-or-place process chain. EVST addresses this by separating image calibration, camera-to-robot calibration, pick-region limits, and result age.
- According to ISO 10218-2:2025, robot-cell integration extends beyond the robot alone. EVST addresses this by treating the feeder, camera, gripper, destinations, safeguards, and exception routes as one application.
- According to the NIST Performance Assessment Framework for Robotic Systems, performance assessment needs measurable procedures and metrics. EVST addresses this by trialing the declared part distribution and recording retries, rejects, grip evidence, and completed transfers.
About the editorial team
The organization supplies industrial and collaborative robot platforms and integrates automation cells for handling, vision, welding, cutting, dispensing, and related operations. Its public engineering guides are intended to help manufacturers define application inputs, interfaces, test evidence, safety functions, and recovery limits before a specific robot or cell is accepted.
The application team can use the declared part set, presentation method, camera geometry, calibration plan, gripper, signal list, and exception routes to prepare an application-specific review. Unknown inputs remain open assumptions until representative testing closes them.
Related engineering resources
- SCARA robot applications for small-parts automation
- Vision-guided robot cell selection guide
- SCARA and vision in a packaging-line application
Frequently asked questions
Is a high vision score enough to allow the pick?
No. The result must also be current, calibrated, inside the approved physical region, compatible with the active part and tool, and associated with a clear exception rule. The score says something about the image result; it does not prove reach, clearance, grip, or destination availability.
How often should hand-eye calibration be repeated?
Set the interval from the application’s stability needs and service events. Always define verification after changes that can move the camera, robot base, tool, fixture, or saved frames. A quick check can run more often than a full recalibration and should block production when its residual exceeds the project limit.
What is the correct response to overlapping parts?
Do not let the detector choose an ambiguous target unless the complete application was designed and validated for that condition. Use mechanical separation, feeder correction, recirculation, a bounded image retry, or a reject path and record the outcome.
When should a vision result expire?
A vision result should expire when part presentation may have changed, the feeder advanced, calibration state changed, a retry altered the scene, or the allowed age elapsed. Store the image and result identity with a timestamp, then require a fresh acquisition instead of moving to a coordinate that may no longer describe the physical scene.