10 Cruise Mapping Engineer (New Grad) Interview Questions (2026)

Outline

Tie to interest in geospatial data, infrastructure problems, or large-scale data systems. If you have GIS, geomatics, or graph-theory coursework, mention it. Show you understand that mapping is upstream of everything — your output is the world model that perception and planning depend on.

Outline

STAR. Mapping data is messy by definition — GPS drift, sensor calibration errors, real-world changes (construction, repainted lane lines). Pick a project where you cleaned or validated messy data. Walk through your methodology.

Outline

STAR. Mapping has this tradeoff constantly — a freshly-detected change might be sensor noise, but waiting too long means the map gets stale. Pick a project (not necessarily mapping) where you balanced similar concerns. Discuss what tipped the decision.

Outline

Brute force: O(N) per query. With a spatial index: O(log N) per query. Walk through KD-tree or grid-based bucketing. For Earth distances, use Haversine formula (or approximate with equirectangular projection for small distances). Discuss accuracy tradeoff.

Outline

Ray casting: shoot a horizontal ray from the point, count edge crossings. Odd = inside, even = outside. O(N) for N polygon vertices. Walk through edge cases: ray passes through a vertex (count carefully), point on the boundary. Mention winding-number as an alternative for self-intersecting polygons.

Outline

Sample both at fixed intervals. Compute pairwise distance (or Frechet distance for trajectory similarity). Threshold to decide match. Walk through Frechet distance — the minimum 'leash length' for two dogs walking along the curves. O(N × M) DP. Discuss tolerance — GPS noise is ~1m, so the threshold has to be generous but not too generous.

Outline

Quadtree or fixed-grid tiling. Each tile stores all map elements within it. For elements crossing tile boundaries: replicate or use boundary handles. Walk through tile size tradeoff (small tiles = more lookups, large tiles = bigger payloads). Discuss versioning — each tile has a version; vehicles request 'tile X version >= Y'. Object storage backing + CDN cache.

Outline

Each vehicle submits localized observations (lane lines, signs, traffic lights). For each map element, collect all observations, compute a confidence-weighted average position. Reject outliers (Z-score or RANSAC). Walk through how to detect a real change vs sensor noise — require N consistent observations from independent vehicles before updating the canonical map. Discuss versioning + rollback if a bad merge ships.

Outline

Compare what the vehicles see (perception output) against what the map says. Recurring discrepancies (multiple vehicles, multiple days) at the same location = the map is likely wrong. Walk through the signal: lane line shifted by 0.5m, missing stop sign, construction zone. Discuss noise filtering (single observation = ignore, 10+ observations = flag).

Outline

Compare both versions against ground truth (high-precision survey data). Compute per-element error (position, orientation). Aggregate to map-level metrics. Walk through how to handle elements that exist in only one version (additions or deletions). Discuss A/B testing in shadow mode — run the new map alongside the old, compare disagreements with downstream perception.