ML System Design: Choose Passenger Drop-off Location

L6 ML design round: when a Waymo passenger arrives at the destination, decide where exactly to pull over. Walk through model features, ranking objective, safety constraints, and how human / regulatory rules interact with the ML score. Requirements At trip end, choose a curbside parking spot for the AV to stop and drop off the passenger. Inputs: ego state + sensor view of the immediate area, map data (lane markings, parking regulations, building entrances), passenger preferences (door choice, accessibility needs), historical drop-off data at the same location. Output: a ranked list of candidate stop poses with scores; the planner consumes the top feasible candidate. Constraints: legality (no-stopping zones, fire hydrants), safety (clear lane visibility, distance from intersections), passenger experience (proximity to the destination door, lighting), and operational considerations (re-merge cost into traffic). Notes Candidate generation. Sample candidate stop poses along the destination block face, e.g. every 0.5 m. Filter for legality from map annotations. Filter for safety from real-time perception (no obstacles in the proposed spot). Scoring. Use a learned ranker that combines (a) walking distance from the stop pose to the destination door, (b) historical success rate at the same micro-location, (c) lane congestion / disruption cost, (d) passenger preference signals. Calibrate the score so the top-K candidates form a usable shortlist for the planner. Hard constraints vs soft scoring. Legality and safety are hard filters — never let the model override them. The learned ranker only orders the remaining feasible set. Long-tail and human override. Rare failure modes (passenger requests an unusual side, construction zones, special events) require either a high-confidence override path or a graceful fallback to a rule-based default. Mention how you'd capture these in training data via labeling pipelines on overridden rides. Evaluation. Offline: walking distance, override rate, ranker-vs-rule agreement. Online: passenger satisfaction (1–5), planner abort rate, time spent searching for a new spot, second-pull-over rate. Run a multi-armed bandit at deployment to balance exploitation of the learned ranker with exploration of new candidate generation strategies. Personalization. Per-rider preferences (curb side, walking budget, accessibility) are stored as features but kept narrow to avoid privacy / fairness issues. Surface the trade-off explicitly. Failure modes to deep-dive. Spot occupied by a parked car at the moment of arrival (replan in <2 s), spot blocked by an oncoming vehicle, no legal spots in the entire block (fall back to the closest legal block and walk). Preparation Sketch the candidate-generation / hard-filter / soft-ranker pipeline before the round; this scaffold lets you defend the ML scope against an L6 interviewer who pushes for 'just write rules.' Pre-derive a calibration story: temperature scaling of the ranker output, isotonic regression against passenger satisfaction. Rehearse the safety / legality discussion as a hard filter, not a soft penalty — collapsing these into the ranker is a common red flag at the L6 bar. Bring 2–3 hardpoints from past projects on long-tail data mining; the interviewer reliably probes whether the candidate has shipped a real ML system into a safety-critical loop.