ADAS radar manufacturing control with AI — final calibration passes the radar, but the result was decided upstream.

The problem

The RF chamber passes the radar — but it cannot see why it calibrated at the limit.

Radar is a safety component: emergency braking and adaptive cruise control depend on it. And in manufacturing, everything converges on a single point: final calibration in the RF chamber, the bottleneck of the line. Every unit goes in, is calibrated against a reference and comes out with a pass or a fail. The problem is that this result is not decided in the chamber — it is decided in everything that came before: the antenna position on the PCB, the housing torque, the radome tolerances.

Three things happen at once on a real radar line, and almost nobody puts them together:

1. The assembly defect is discovered late and expensively — an antenna rotated half a degree or a housing closed with uneven torque is invisible once the unit is sealed. It shows up as a poor calibration in the chamber, when the radar is already finished and rework costs the most.
2. Today's "just barely passed" is tomorrow's field failure — a radar that calibrates hard against the limit can drift out of calibration with vibration and thermal cycling. A binary pass/fail cannot tell a robust unit from a marginal one, and process drift (calibrations creeping ever closer to the limit) triggers no alarm at all until the rejects start.
3. Safety traceability demands what the line does not keep together — reconstructing the complete process of every unit: assembly, torques, radome batch, calibration curve. When the audit or a field problem arrives, the file has to be stitched together by hand, system by system, and containment ends up covering the entire batch.

The outcome is always the same: the calibration chamber acts as the only barrier, rejects appear when the drift has already been running for hours, and the knowledge of why this part number calibrates worse after a batch change lives in the head of the line technician. The day they move to another shift or another company, it leaves with them.

How it fits the IRIS system

iLEAN does not replace the calibration chamber — it joins what the line keeps apart.

The ADAS radar problem is not a lack of equipment: the line already has vision at some stations, nutrunners that log torque and an RF chamber that measures precisely. The problem is that each piece of data lives on its own island and the calibration result arrives without the context that explains it. iLEAN acts as the putty that joins assembly vision, torques, radome batch and the calibration curve of every unit, without asking you to change the chamber or the layout of the line.

Edge verifies assembly on 100% of units before closing. Connect stitches every calibration to the serial number. JIDOKA AI detects drift before the first reject. Agents keep the file for every unit. The person decides what to correct.

The four iLEAN pieces applied to ADAS radar manufacturing control:

• Edge — vision at the station before closing: antenna position and rotation on the PCB, housing seating and torque, radome condition. On 100% of units, not by sampling, because the assembly defect that slips through today is the poor calibration twenty minutes from now. It runs locally: if the plant loses the network, Edge keeps verifying and recording. What is critical does not depend on WiFi.
• Connect — integration with the RF calibration chamber and the rest of the stations: it stitches every calibration curve (power, phase, pattern by angle) to the unit's serial number, together with what vision saw and the torques applied. Pass/fail stops being an isolated data point and gains the full context of the process.
• JIDOKA AI — analyzes the distribution of calibrations continuously and compares it with the line's history for that part number. It detects process drift — calibrations creeping toward the limit, all of them passing — and flags it before the rejects start, pointing to what changed upstream. It does not act alone: it proposes, the line lead decides.
• Agents — maintain the file for every unit: verified assembly, torques, batches, calibration curve, reference used. Ready for the safety audit, and ready to narrow down the affected population by real process window when a problem appears in the field.

See the full IRIS architecture →

Before and after

Classic ADAS radar line vs. control with iLEAN

Aspect	Final calibration as the only barrier	With iLEAN Edge + Connect + JIDOKA AI + Agents
Assembly verification before closing	Sampling, or trust in the process	Edge vision on 100% of units
Calibration result	Isolated pass/fail in the chamber	Curve stitched to the serial with process context
Process drift	Invisible until the first reject	Detected as the population shifts toward the limit
Cause of a poor calibration	Unit sealed, cause invisible	Traced to antenna, torque or radome batch
Per-unit file for safety	Manual reconstruction, system by system	Automatic, by serial number
Containment after a field problem	Whole batches, as a precaution	Population bounded by real process window

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• ADAS radar manufacturing line with final calibration in an RF chamber as the bottleneck, partial or sampling-based assembly verification before closing, and calibration rejects documented over recent months.
• Pilot: Edge with vision at the station before closing + Connect integrating the calibration chamber. JIDOKA AI learns the line's distribution over 4-6 weeks before it starts anticipating. First value expected within a few weeks: the curve-to-serial stitching and the detection of the first drifts show up early, before the model is fully tuned.
• Expected reduction in calibration rejects of ≥30% by attacking drift upstream — estimate to be validated with the real history of your chamber.
• Indicative payback between 5 and 12 months. The hard lever: every unit that does not reject in calibration frees up the bottleneck of the line, and reworking a sealed radar is among the most expensive operations in the process.
• A recurring benefit that does not enter the ROI but carries weight: the per-unit file is ready for the safety audit, and containment after a field problem goes from whole batches to populations bounded by process window.

And the quality manager's reasonable doubt

"What if the AI flags a drift that isn't there and we stop the line for nothing?" — JIDOKA AI does not act on the line by itself. It proposes; the line lead decides; the correction is applied or it is not. Hallucination is a problem of free generation, not of anchored tasks: in tasks where the AI cross-references calibration curves with history and assembly data, the best models brought error below 1.5%^[1]. And even so, what is critical goes to the safety rings — JIDOKA AI lives in the outer ring, proposes inward, and any action on the process is signed by a person. Never the other way round.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.