NADIR / Technology / 01 of 04

Every sensor is wrong. The work is knowing by how much.

A camera, a radar, a lidar and an inertial unit each describe the same stretch of road in different units, with different failure modes. Post-repair drift is what happens when one of them quietly starts describing a road the others do not recognise.

Geometry

The cost of a degree.

Extrinsic calibration is the transform between where a sensor is physically bolted and where the vehicle believes it is pointing. Six numbers: three rotations, three translations. The rotations are the dangerous ones.

A camera rotated nine tenths of a degree in yaw still produces a sharp, correctly exposed, perfectly focused image. Nothing in the frame looks wrong, and no self-test inside the camera has any reason to complain. But project that error down the road to automatic-braking range and the vehicle places an object 1.57 metres from where the object actually is. That is wider than a lane line. It is the difference between braking for a pedestrian and steering past one.

This is the entire argument for monitoring after the fact. The failure is not in the image. It is in the geometry, and geometry only shows itself when you compare one sensor against another.

Extrinsic drift, camera to vehicle A 0.90° yaw error projects to 1.57 m of lateral error at 100 m — wider than a lane line, and invisible in the image itself.

Modalities

Four instruments, four ways to be wrong.

Each modality fails in a way the others do not, which is exactly why disagreement between them carries information. A camera's intrinsics drift slowly and surface as reprojection residual — sub-pixel near the optical centre, several pixels out at the corners where the distortion model does most of its work. Stereo depth error grows with the square of range, so a fifth of a pixel of rectification bias is invisible at ten metres and ruinous at a hundred.

Radar is immune to most of the optical problems and has its own: a boresight rotation shifts the entire range-Doppler field, and the static world stops sitting where ego-velocity says it should. Lidar hands you geometry directly, which makes a pitched mount obvious in the ground-plane residual and close to invisible everywhere else.

Camera intrinsics Reprojection residual across the image plane. Corner error runs an order of magnitude above centre error, because that is where the distortion model carries the most weight.

Stereo depth error Depth error against range. A 0.15 px rectification bias is systematic rather than random, so it does not average away across frames.

Range-Doppler field Returns that should sit on the static-world gate have moved off it, and one target is rejected as clutter by cross-modal conflict.

Ground-plane fit A plane fitted through the lidar ground return. A 0.18° pitch bias appears as residual that grows with range and exceeds tolerance at 25 m.

Attribution

The slow movers.

Not all drift arrives with the repair, and separating the two is most of what makes post-repair scoring hard. Gyro bias walks across a drive and is bounded by the Allan deviation floor of the part rather than by anything a technician did. Wheel odometry diverges from GNSS by a percentage of rolling radius that changes when the tyres change. A camera bracket expands with temperature at a rate repeatable enough to model and large enough to matter — better than half a degree between a cold morning and a hot afternoon.

NADIR anchors every score to the repair-order close timestamp for exactly this reason. A bracket warming in the sun is not a bracket that was installed crooked, and a system that cannot tell the two apart will spend its first month crying wolf.

In-run gyro bias Bias walk against the Allan deviation floor of the part. This sensor is inside spec — the drift has to be modelled, not flagged.

Odometry against GNSS 3.2 m of divergence over 212 m is a rolling-radius error. It changes with tyres, not with calibration.

Thermal expansion 0.0104°/K is 0.55° across the working range — 0.97 m of lateral error at 100 m from weather alone.

None of this is detectable from a single sensor. A camera cannot tell you it is pointing nine tenths of a degree off, because from inside the camera nothing has changed.

Only disagreement between sensors that ought to agree carries the signal, which is why every number on this page is a residual rather than a reading.

Technology

The rest of the category.

01Sensing and driftWhat physically moves after a repair 02Calibration geometryWhat a bay measures, and what it misses 03Scoring methodResidual to tier, with calibrated bounds 04Boundary and evidenceShadow mode and signed lineage

Want this pointed at your fleet?

We are onboarding fleets and repair networks a few at a time. Tell us what you run and we will reach out.

Join the waitlist