“Can a robot pick randomly oriented parts without stopping the conveyor?” The answer for a Thai factory depends on much more than camera recognition. It depends on the time between observation and contact, changing belt speed, grip reliability, downstream insertion or tightening, and a safe restart after a stop. This guide turns 3D vision conveyor tracking into a procurement decision: define the work, ask comparable questions in the RFP, run a 90-day proof of concept, and accept the cell through FAT and SAT.
What Nikon announced in September 2026—and what it did not guarantee
On 30 September 2026, Nikon announced a 3D vision tracking feature for its robot vision system, scheduled for release in October 2026. Nikon says existing customers can obtain the feature through a paid software update. Its 3D cameras measure the position, inclination, attitude and orientation of workpieces moving on a conveyor or other transport equipment. Nikon cites randomly oriented castings and parts for picking, grommet insertion into vehicle bodies, and bolt tightening on moving workpieces. The company says its particular system requires no additional equipment such as an encoder. That claim describes Nikon’s architecture; it does not make encoders unnecessary for every tracking design. Source: Nikon announcement, 30 September 2026.
Nikon’s November 2025 release described a compact model dedicated to 2D vision tracking. That creates a useful decision boundary. If part height and attitude change the contact point or tool angle, evaluate 3D. If the part is flat, singulated and held at a repeatable height, 2D may be sufficient. Neither release gives a guaranteed throughput, accuracy or uptime for your Thai line, and neither proves that an installed robot, gripper, safety system and camera can be upgraded without engineering work. Treat the announcement as a new option, then set your own acceptance values.
Define the moving task before choosing a camera
Put one sentence at the top of the RFP: “Pick randomly oriented, single-layer castings from a running conveyor and place them in a fixture,” or “insert a grommet into a moving body opening,” or “tighten a specified point on a moving sheet-metal part.” Name the part, movement, robot action and good-part criterion. That sentence stops a quotation from becoming a contest of camera pixel counts.
Picking a stationary part from a bin and following a moving part are different timing problems. Bin picking concentrates on occlusion, collision-free paths and grasp candidates. Conveyor tracking also has to account for the time stamp of an observation, motion between camera and robot, speed changes, and whether the part is still reachable when the tool arrives. Our earlier robot vision cost and ROI guide covers general supply, precision and lighting issues. This guide narrows the scope to dynamic hand-off and acceptance. The picking robot implementation guide helps compare stopped and moving picks.
When 3D adds value, and when 2D may be enough
3D matters when variation in height or tilt changes the robot’s approach. Examples include a casting that can arrive on either face, a molded part that rotates on the belt, or a moving body surface that changes the axis of an insertion or tightening tool. Before writing “3D required,” measure height distribution, acceptable attitude and tool-angle tolerance. Singulation or simple fixturing may lower total project cost more than adding a sophisticated vision stack.
“3D vision” bundles several functions: depth capture, recognition, pose estimation, time stamping, tracking, coordinate hand-off and robot motion. Reflective metal, black resin, deep recesses, oil, dust and variable factory lighting can affect the visible surface. Do not test only pristine samples. Include the largest and smallest parts, finish variations, wear, contamination, partial occlusion and mixed SKUs. Record which samples are intentionally excluded and how they will be diverted.
Compare architectures against the same task. Nikon says its new feature does not require an additional encoder. ABB’s Conveyor Tracking Module describes connection ports for cameras, conveyors and encoders. These are distinct product approaches. Ask who estimates part position at the future contact time, how belt slip is handled, and what happens after a stop or restart. FANUC presents iRPickTool with 3D iRVision as a line-tracking and part-detection option; its dynamic applications article gives further context. Product lists do not replace a trial with the same workpieces and pass criteria.
Data the RFP should provide
A CAD file is not a record of how parts actually arrive. Give bidders videos across shifts, the SKU mix, normal/minimum/maximum belt speed, acceleration and short-stop logs, good and defective samples, and a drawing of the reachable work area. If video cannot leave the plant, state how the supplier will observe and test on site.
| Input | Data to supply | Design consequence |
|---|---|---|
| Workpiece | Material, tolerances, mass, finish, prohibited grip surfaces, SKU count | Camera, illumination, gripper and recognition |
| Supply | Single layer or overlap, orientation distribution, spacing, bounce and mixed parts | Field of view, recapture and singulation |
| Conveyor | Speed range, acceleration, belt slip, stop and restart | Tracking method and robot reach |
| Process | Pick, insertion, tightening, inspection and rework sequence | Tooling, force control and downstream signals |
| Controls | PLC, I/O, network, existing encoder and clock synchronization | Integration effort and ownership |
| Environment | Lighting, oil, dust, vibration, cleaning and maintenance skills | Enclosure, calibration and support |
| Safety | Human access, guarding, doors and emergency-stop reset | Risk assessment and safety devices |
Specify what happens to an unrecognized part. Can it continue downstream, be diverted, or must the conveyor stop? State the response to a double pick, failed grip, dropped part and wrong SKU. A recognition percentage alone does not answer these questions. If an operator collects rejects, include that work and the safe access route in total cost.
Calculate the time window before asking for speed claims
Between the camera view and physical contact are exposure, image transfer, 3D reconstruction, object identification, trajectory update, robot approach and gripper action. The average delay is not enough; examine its upper tail. A part that was reachable when photographed may have passed the safe pick zone by the time the robot arrives.
Under a constant-speed approximation, use measured values in a simple planning relationship: available time equals the distance from capture to the last safe contact position divided by part speed. With acceleration or deceleration, use the measured speed profile to estimate position over time and verify the actual tracking window. Compare that with the sum of capture and processing, coordinate communication, trajectory correction, robot approach, grip confirmation and a safety margin. This is a design worksheet, not a vendor performance formula. If the margin is too small, moving the camera upstream, increasing spacing, changing the gripper or altering the robot position may matter more than a faster camera.
For illustration only, assume a belt speed of 0.5 m/s and 0.8 m from camera to the end of the safe pick zone. The available window is 1.6 seconds. This is an example calculation, not a typical Thai factory speed. Subtract approach, confirmation and safety time before allocating time to image processing. Repeat the calculation under acceleration, deceleration, manual jog and restart.
Measure good output, not just camera recognition
A camera can identify a workpiece that the robot cannot reach. Define a countable chain: parts fed, candidates detected, correct identity and pose, targets inside the safe reach zone, successful picks, completed insertion or tightening, final inspection passes, and human interventions. Log all counts by batch with timestamps and reason codes. The chain helps the camera supplier, robot integrator and plant find where a loss occurs.
Give every rate a denominator. For example, pick success might mean confirmed grips divided by issued pick commands, while feed-to-good yield uses all eligible parts fed. Excluded parts need explicit reason codes. Review short windows near the worst observed condition as well as long continuous runs. Acceptance thresholds depend on part value, safety and rework options; no universal percentage should be copied into an RFP.
Tightening quality includes torque, angle, thread damage and tool slip. Insertion includes force, seating depth, orientation and surface damage. Picking includes dropped parts, collisions and manual rework. The scope of measurement instruments and lot traceability belongs in the quote.
A 90-day PoC with four decision gates
Ninety days is an example planning horizon, not a standard installation duration. Existing robots, safety equipment and sample availability may shorten or extend it. Give each stage a deliverable and a stop decision.
| Example period | Gate | Evidence | Decision |
|---|---|---|---|
| Days 1–15 | Baseline | Workpiece, conveyor and safety data; exclusions; KPIs | Defer purchase if actual data is missing |
| Days 16–35 | Imaging and grip | Samples for every SKU, lighting plan, grip candidates, failure causes | Redesign feeding if difficult parts dominate |
| Days 36–65 | Dynamic cell | Speed variation, timestamped commands and motion, reject handling | Change layout if timing margin is inadequate |
| Days 66–90 | Plant-like run | Continuous operation, changeover, stop/restart, quality and maintenance | Defer production investment if plant thresholds fail |
Observe the process before simplifying it. Operators may quietly orient parts, create spacing or separate defects. Removing that manual step changes the arrival distribution. The PoC must reproduce the future process, not an orderly demonstration. Include night-shift lighting and conditions near a cleaning interval.
Dynamic tests must include acceleration and intermittent stops, not only a smooth constant-speed belt. Check that an old target coordinate cannot be reused after a safety gate opens, a network link drops, or the conveyor restarts. Decide which lens-contamination condition creates an alarm. Record unsuccessful samples, timestamps, root causes, corrective actions and residual risks alongside successful video. Give Thai plant staff usable English or Thai screens and calibration procedures.
Compare the whole cell quotation
Break out camera, lens, light, control PC, software license, robot, end effector, force/torque devices, conveyor changes, PLC connection, guarding, fixture, engineering, installation, commissioning, training, spares and local support. Include software renewal, additional SKU learning, recalibration and night-shift response in ownership cost.
Nikon’s paid update for existing customers is a possible procurement route. Its announcement alone does not establish installed-hardware compatibility, Thai service availability or price. Compare updating an existing cell, a different tracking module, 2D vision with better part presentation, and a brief stop-and-pick operation against the same output requirement. Ask for prices and schedule effects when SKU count rises, belt speed changes, a gripper is replaced, or the camera must move.
Separate FAT from SAT
Factory Acceptance Testing checks agreed workpieces and repeatable conditions at the supplier or test facility. Site Acceptance Testing checks the actual Thai line: light, vibration, conveyor slip, PLC timing, guards and people. Passing FAT does not establish SAT performance.
| Test | FAT | SAT addition | Evidence |
|---|---|---|---|
| Recognition and pose | All SKUs, good and bad samples, occlusion, light variation | Actual light, dirt, vibration and mix | Source image, decision and reason |
| Dynamic tracking | Speed range, acceleration, stop/restart | Actual belt slip and PLC time offset | Timestamped events |
| Process quality | Pick, insertion and tightening targets | Downstream inspection, rework, long run | Measurements and lot ID |
| Safety | Guard, gate, emergency stop and reset | Actual access and maintenance tasks | Risk assessment and test record |
| Operation | Alarm, recipe change and calibration | Local operator performs the procedure | Manuals and sign-off |
Put sample counts, exclusions, thresholds, retest limits, log location and approvers in the contract. If an FAT issue is carried into SAT, list it with a temporary control and a condition for production release. ISO 10218-2:2025 addresses safety requirements for industrial robot applications and cells. A robot catalogue or the existence of the standard does not prove that this integrated cell is safe. Include dropped parts, moving-tool contact, human entry and maintenance restart in the cell risk assessment.
Set the responsibility boundary for maintenance in Thailand
When the camera vendor, robot maker, gripper maker, conveyor supplier, integrator and plant maintenance team are different companies, a failure can become “we sent the coordinates” versus “the robot followed the command.” The interface table must define clock reference, coordinate system, target identity, timeout, error response and reset authority, not just I/O names. Agree who diagnoses and fixes faults, what is included in warranty, and when support becomes billable.
Specify screen, alarm and procedure languages for the people who will use them. Test actual operators in SAT. If only the supplier can recalibrate the camera or add an SKU, document price, lead time and remote-access permissions. A reliable cell should have an explicit local recovery procedure after power loss, network failure and component replacement. Our robot hand and gripper selection guide helps isolate grip force and finger geometry from vision performance.
Pre-purchase site checklist
Use measured values with measurement date and owner. Distinguish conveyor set speed from measured part speed, and include start-up after a changeover as well as peak steady speed. Record orientations in actual feed order across shifts. Compare daylight and night shifts, pre- and post-cleaning, and lighting from adjacent equipment. Mark both theoretical robot reach and the smaller area where safe deceleration and retreat remain possible. Follow dropped parts to the point where they are recovered; guard entry for cleanup is a real maintenance and downtime cost.
Set a common clock for camera, PLC and robot logs. The RFP should state log retention duration, how events link to lot IDs, and whether and for how long source video may be retained under plant rules. A timestamped fault cannot be investigated reliably when the three devices use different clocks. Build a baseline by SKU and shift: fed pieces, good pieces, misses, rework and stop minutes. Specify whether throughput uses net running time or the entire shift, and whether upstream shortages count as cell downtime. Use the same definitions before the PoC and during FAT/SAT. A favorable demonstration is not evidence of improvement if the denominator changed.
Promote abnormal events and changeovers into tests. Parts can turn, slip, collide, overlap or bounce. The system need not recover every case, but it must reject unsupported cases safely and record why. For a recipe change, verify that camera exposure, lighting, gripper, robot trajectory and inspection rules all follow the same SKU ID. Specify how the line handles parts already between camera and robot when a recipe changes, and whether a failed changeover requires an empty line before restart.
Put achieved values, unmet conditions and the extra cost of corrections on the same decision page. Get manufacturing, quality, maintenance and safety approvals, and record test date, software and recipe versions, sample lot and instrument calibration in the signed acceptance record. Define the warranty envelope by speed, SKU, pose, lighting, contamination and continuous-run duration. Include glossy, dark, worn, contaminated and partly hidden samples within drawing tolerance. Identify FAT samples by lot or image and add an independent lot in SAT. If a test passes only after excluding a condition, manufacturing, quality, maintenance and safety owners must confirm that the production line can actually enforce that exclusion. Otherwise redesign the feed, gripper, layout or process instead of silently weakening the threshold.
Frequently asked questions
Does 3D vision conveyor tracking eliminate the encoder?
No universal answer exists. Nikon says its announced feature needs no additional encoder; other designs include one. Evaluate belt slip, speed variation, distance between capture and contact, and the evidence you need from each candidate.
When should we choose 2D rather than 3D?
Use the variation in part height and attitude to decide. A flat part at a controlled height may need only planar correction. Keep a feed-improvement option in the comparison before paying for more vision capability.
How many parts per minute can we guarantee?
Do not derive a plant guarantee from a product release. Throughput depends on part mix, spacing, camera-to-pick distance, robot path, grip time and reject handling. Measure your actual pattern in the PoC and use the distribution and continuous-run result for FAT/SAT.
Will an existing Nikon system need only a software update?
Nikon offers the new feature as a paid update to existing customers, but compatibility of specific cameras, robots, tools, PLCs and safety equipment—and the work required on site—must be confirmed individually. Quote the update and modifications separately.
Why can a PoC pass while production fails?
A demonstration may cover too few SKUs, speeds, lighting states, contamination levels and stop/restart events. Test the real arrival distribution, preserve failure logs, and repeat the agreed KPIs on the actual line at SAT.
Conclusion
The purchasing question is whether the cell can measure the pose of a moving part, predict where it will be at contact, and complete the pick, insertion or tightening safely and to specification. Nikon’s September 2026 announcement expands the choices; it does not determine your line’s quality or capacity. Fix inputs and responsibility in the RFP, reduce uncertainty in the PoC, then carry the same KPIs through FAT and SAT.
If your Thai plant is considering a moving-part pick, insertion or tightening process, contact TOMAS TECH even before all drawings and takt data are ready. We can help define the target operation and the site data needed for a useful RFP.
