Executive summary — the one-minute version
|
🏢 A real station, verified end to end An 11-step manual assembly SOP — connector placement, multiple screw groups, torque-judgment steps — executed under continuous camera supervision. In the recorded end-of-flow verdict: 5 of 5 detections OK, 100.00% OK rate, green PASS, one missed-screw timeout caught in real time. |
🕯️ One cable per camera Each GMSL2 camera (1–3 per station, chosen by target type and resolution) feeds the edge box over a single FAKRA coaxial cable — uncompressed video and 12 V power in the same cable, up to 15 m — so the cell stays wiring-simple, and >100 dB HDR keeps the image usable under harsh station lighting. |
|
🧩 Built low-code, not from scratch The application runs on uGrab, a low-code industrial AI vision platform: a visual node workflow with an out-of-the-box YOLO detection and OpenCV-Sharp toolkit. SOP logic, timeout supervision and records are configuration — not a bespoke C++ project. |
🔑 Traceability by design Every unit is keyed by a scanned serial number. Step times, detection results, warnings and the final PASS/NG verdict are written to a per-SN record for SPC — with API hooks for MES/QMS integration (the deployment pre-checks every scanned SN against a customer-side API). |
Walk down any manual assembly line and ask one question: how do you know every unit was built exactly to the standard operating procedure? For most factories the honest answer is paperwork, spot checks and end-of-line inspection — all of which discover mistakes after they have already been made. This article is the story of a station where that question is answered in real time, by a camera, on the edge: an AI-enforced SOP — what we call an AI-SOP — deployed on EyeCloud AI hardware, with the evidence shown below.
Human error remains the dominant failure mode of manual assembly. Industry surveys commonly attribute around 80% of quality defects to human factors, and a widely cited global study of 100+ manufacturers found human error behind roughly 23% of unplanned downtime. The cost of poor quality is estimated at 15–20% of sales revenue for typical manufacturers — which is why connected-worker and operator-guidance platforms are one of the fastest-growing segments in manufacturing software, projected to grow from about USD 8.6 billion in 2025 to over USD 20 billion by 2030 (18.5% CAGR).
Traditional poka-yoke — hard fixtures, sensor pins, interlocks — catches only the errors an engineer predicted. An AI-SOP does something structurally different: a camera watches the station, a vision model detects each step’s outcome, and software enforces the sequence. Wrong part, missing screw, skipped step, out-of-sequence action — each becomes a flagged event the moment it happens, not a defect found at final inspection.
Definition. AI-SOP (AI-enforced standard operating procedure): a computer-vision application that verifies each step of a manual work instruction in real time — guiding the operator, warning on deviations, and recording an auditable per-unit result. It is vision-based error proofing (digital poka-yoke) and SOP compliance monitoring in one system.
The deployment is a mixed manual-assembly station: operators place connector groups, drive multiple screw groups, and perform torque-judgment steps on a precision electromechanical assembly. The SOP runs to eleven supervised steps per unit. The manufacturer’s requirements were concrete:
| Layer | Deployed |
|---|---|
| Capture | 1–3 overhead GMSL2 cameras — the count set by the targets: one wide view covers a full bench at connector-level detail, while small screw features, occluded or multi-face stations take 2–3 cameras. One FAKRA coax per camera (lossless video + 12 V power over the same cable). Platform: SDC800GY-C C-Mount camera (4M/5M/8M switchable @ 30 fps, >100 dB HDR, YUV output) or the GMSL2/3 M12 family (SDC300/SDC500/SDC800). |
| Edge AI compute | RobooPi P53NVNA AI computing box: NVIDIA Jetson Orin Nano 8 GB (up to 40 TOPS), Ubuntu 20.04, four FAKRA camera inputs with hardware frame sync, 128 GB storage, HDMI for the line display, GPIO/I2C for station I/O. |
| Vision software | uGrab low-code platform: visual node workflow, YOLO object detection and OpenCV-Sharp toolkit, SOP step engine with timeout supervision, per-SN SPC records and API hooks. |
| Operator interface | Step checklist with live check-off and step timers, on-screen prompts of the next action and target location, warning banners, and a station-level PASS/NG verdict per unit. |
| Peripherals & station I/O | Fill light for stable illumination at the verify points; speaker for voice prompts and rejection alarms; DI/DO control module to interlock with the station — stack light, reject gate or PLC handshake — driven from the box’s GPIO/I2C and audio interfaces under uGrab node control. |
| Records & integration | Per-SN traceability record (steps, detections, warnings, verdict); SN pre-check against the customer’s API before the unit is accepted; data export for SPC/MES/QMS. |
The recorded session shows the full lifecycle of one unit:
Mid-session, the operator left one screw group idle past its time window. The station did not wait for end-of-line inspection:
That single behavior — catching a skipped screw group at the station instead of at final test, or worse, in the field — is the entire economic argument for AI-SOP.
Every claim below is read directly from the recorded deployment session:
| Metric (observed in the recorded session) | Result |
|---|---|
| SOP steps under camera supervision | 11 mixed steps — connector placement, screw groups, torque-judgment steps |
| Detection confidence range (connectors, screws) | 0.89 – 0.97 across the session |
| End-of-flow verdict | 5 total detections, 5 OK, 0 NG — 100.00% detection OK rate, PASS |
| Missed-step response | Timeout warning banner + NG indicator in real time (step |
| Traceability record per unit | Scanned SN, per-step times, detections with confidence, warnings, final verdict; SN pre-checked against customer API |
The station’s capture layer borrows deliberately from automotive-grade vision. GMSL2 (Gigabit Multimedia Serial Link 2) serializers push uncompressed video over a single coaxial cable with FAKRA connectors — the same signaling cars use for surround-view cameras, engineered for vibration, EMI and long cable runs:
| Consideration | GMSL2 camera (SDC800GY-C class) | Typical USB camera |
|---|---|---|
| Cabling | One coax: video + 12 V power over the same cable, up to 15 m, locking FAKRA connector | USB 3.0 limit ~3–5 m without active extension; separate power; connectors not rated for vibration |
| Latency | Uncompressed SerDes link — microsecond-class transport latency, so step verdicts land inside takt-time budgets and operator feedback stays immediate | Compressed USB/RTSP pipelines add encode/decode buffering — tens to hundreds of milliseconds of glass-to-glass delay |
| Image fidelity | Pixels cross the link uncompressed: the model judges exactly what the sensor captured — no codec artifacts softening screw threads or small edges | Codecs trade image truth for bandwidth; compression artifacts can erase the fine detail small-feature detection depends on |
| Image robustness | Automotive-grade imaging chain; >100 dB HDR for backlit/glare-heavy stations; SNR 39.12 dB; stable in ≤800 ms from power-on | Commodity sensors; exposure tuning per model; HDR rare |
| Host load | YUV (UYVY) out of the camera’s ISP — no debayering on the Jetson; cycles stay free for inference | Raw sensor formats common; debayer/ISP load lands on the host |
| Scaling the cell | Up to 4 FAKRA channels per RobooPi box with hardware frame sync (GMSL2/3 and FPD-Link deserializer options: MAX9296 / MAX96792 / DS90UB954 / DS90UB9702) | Multi-camera USB is host-controller-bound; frame sync across USB cameras is impractical |
| Environment | Camera rated −20 °C to +65 °C; GMSL2/3 M12 variants target automotive and high-temperature industrial use | Consumer temperature ranges; no automotive pedigree |
Industrial machine-vision cameras and automotive cameras used to force a choice. GigE/USB machine-vision cameras give you the full C-Mount lens ecosystem but put a bulky camera and a thick cable at the point of use; automotive GMSL cameras give you the rugged remote head and the single coax but fix the lens. The SDC800GY-C deliberately combines the two: a 30 × 30 × 35.4 mm sensor head that mounts where the work happens, an automotive SerDes link that carries lossless video and 12 V power back to the cabinet over one thin coax, and a standard C-Mount ring that accepts the industrial lens ecosystem — FOV60 for a full-bench view, FOV35 for tighter detail, or a macro to resolve a screw thread. Lens choice becomes a per-station decision made after you know the part — not a constraint baked into the camera.
All inference and SOP logic run on the RobooPi P53NVNA at the station — no video leaves the cell except records and stills:
uGrab is EyeCloud AI’s low-code industrial AI vision platform. The AI-SOP application is assembled from a small set of building blocks — which is exactly what keeps a rollout measured in weeks:
The same graph pattern re-appears across stations: only the step definitions, the trained detection model and the layout overlays change. That is the practical meaning of low-code on a production line — the second station costs a fraction of the first, and your own engineers can own the graph after training.
| # | Step | What happens |
|---|---|---|
| 1 | Scope the SOP | Walk the work instruction; decide which steps are vision-verifiable, which need API/IO gating, which stay manual; define time windows and the warning policy. |
| 2 | Mount & light | Mount 1–3 GMSL2 camera heads overhead — positioned per verified step location, C-Mount lens picked per FOV; add a fill light where illumination varies; one coax per camera to the RobooPi box. |
| 3 | Collect & train | Capture part states at the station (OK, missing, wrong-position); label and train the YOLO model; validate to the confidence band your process requires. |
| 4 | Build the SOP graph | In uGrab: bind steps to detections, set prompts/overlays and timers, wire API pre-check for SN and any tool signals. |
| 5 | Pilot with the line | Run side-by-side with the paper SOP; tune windows and warnings with operators; confirm the verdict record lands in SPC/MES. |
| 6 | Roll out | Clone the graph to sibling stations, swap models/overlays per variant; monitor per-step times for cycle-time and training insights. |
AI-SOP verification is the use of a camera and on-edge AI models to enforce a standard operating procedure in real time: each manual step — placing a part, driving screws, confirming torque — is verified in the live video, operators are guided and warned on screen, and every unit gets an auditable PASS/NG record keyed to its serial number.
A fixed camera views the station; an object-detection model (YOLO class) is trained on the station’s part states (present, missing, wrong position). The SOP engine binds each work-instruction step to a detection condition and a time window. When the detection confirms the step, the checklist advances; when a window expires without confirmation, the system warns and flags NG. Nothing is reconstructed after the fact — verification happens while the operator works.
Between one and three GMSL2 cameras per station — the count follows the targets: one wide view can cover a full bench for connector-level checks, while small screw features, occluded or multi-face stations take 2–3 cameras (e.g. the SDC800GY-C C-Mount: 4/5/8MP modes, >100 dB HDR, YUV output). Plus one edge AI box (RobooPi P53NVNA: Jetson Orin Nano, 40 TOPS, four FAKRA inputs) per cell, a display for the operator, and where useful a fill light, a speaker and a DI/DO module for station interlocks. No PC in the cell; one coax carries video and power per camera.
Three practical reasons: reach and reliability (one locking FAKRA coax up to 15 m with power included, vs. ~3–5 m practical USB 3.0 and fragile connectors); image robustness (automotive-grade sensor/ISP with >100 dB HDR for glare-heavy stations); and host efficiency (YUV output means the Jetson spends its cycles on inference, not debayering). GMSL2 also scales cleanly: up to four hardware-frame-synced channels per box.
Yes — that behavior is shown in this case study: a screw group left idle past its time window triggered a red “timeout, not executed” warning and an NG indicator at the station, in real time. The detection model catches visually confirmable states; steps that vision cannot prove (e.g. torque) are gated through the step engine via API/IO inputs.
The operator scans the unit’s serial number; the software pre-checks the SN against a customer-side API, then records every step time, detection (with confidence), warning and the final verdict under that SN. The record is available for SPC and export to MES/QMS — so a quality engineer can replay any unit’s build from the data.
The platform exposes API hooks that the deployment already uses for SN pre-checking; step results, verdicts and per-SN records can be pushed or pulled through the same integration layer. If your MES accepts HTTP API calls or file/CSV exports, the record flow is straightforward.
The station graph itself is low-code configuration, so effort concentrates where it should: scoping the SOP, collecting training data at the station, and validating detection confidence. As a planning figure, pilot stations of this class go live in weeks, not quarters — and sibling stations reuse the same graph with a new model and overlays, which is where the low-code approach pays back.
If manual assembly errors, paper SOPs or thin traceability are on your quality agenda, we can replicate this playbook on your station — camera selection, edge compute, the uGrab application graph and MES integration. Explore our cameras and edge AI products, browse our engineering case studies, read the uGrab low-code vision platform overview, or contact our team to scope your first AI-SOP station.
