Autonomous vehicles
Stress-testing perception stacks against scenarios that don't happen often enough
Northwind used synthetic LiDAR + camera scenes to systematically cover edge cases their fleet rarely encounters in the real world.
−38%
critical-scenario miss rate
142
edge-case scenarios in CI
8M
labelled synthetic frames
The challenge
Northwind's perception team needed to validate their model against night-time pedestrian crossings, partially-occluded cyclists, and adverse weather. Real fleet data contained these scenarios at <0.4% frequency, making targeted regression testing impossible.
How they used Epineone
- Defined a scenario catalogue of 142 long-tail driving situations with parameterised lighting, weather, occlusion and trajectory inputs.
- Generated 8M synthetic frames with paired LiDAR point clouds, RGB camera, and ground-truth 3D bounding boxes.
- Wired synthetic scenarios into the existing CI pipeline so each model release is graded against the full catalogue automatically.
- Used Epineone's bias-parity tooling to ensure performance was consistent across pedestrian appearance categories.
"Epineone gave us a regression suite for the scenarios we hope never happen — which is exactly the suite that matters for safety."