Robot Human-Avoidance — Open-Ended Design

Robotics-team intern coding slot that is really an architecture / pseudocode discussion: design how a warehouse robot detects and avoids humans, picking sensors and data structures along the way. Requirements Open-ended prompt: "How would you make this robot avoid people?" Interviewer expects pseudocode plus an architecture sketch in plain text, not a compiled solution. Drives the candidate through three explicit decision axes: Sensor selection (cameras, depth / LiDAR, ultrasonics) and the latency vs accuracy trade-off. Data structure for the local world model — occupancy grid vs object list vs trajectory buffer. Decision loop — perception → planning → actuation, including how often each tier runs. The interviewer in this loop actively suggested options when the candidate stalled; treat it as a guided design dialogue, not a silent whiteboard. Notes Robotics SDE intern loops at Amazon are trending toward this open-design format; the same recruiter screen that promised "BQ + LeetCode" delivered architecture instead, so do not assume the round will be conventional coding. Strong answers walk through a layered pipeline: low-frequency global planner, mid-frequency local planner with dynamic obstacles, high-frequency reactive controller. Name each loop's hertz target. A small but explicit safety conversation scores well — fail-stop behavior when sensors disagree, a watchdog that halts motion on perception drop-out. Bring up evaluation: how do you tell offline that the avoider works? Replaying logged human trajectories against the policy is the standard answer. Preparation Sketch a one-page reference architecture (perception → world model → planner → controller) and rehearse narrating it in under five minutes. Read one published reference on dynamic-obstacle avoidance (the Probabilistic Roadmap + velocity-obstacle family) so the trade-off vocabulary is at your fingertips. Have a fallback story for sensor failure ready — interviewers in this slot reliably ask "what if the depth camera dies mid-run?" Practice the dialogue style: pause to ask the interviewer which axis they want depth on (sensors vs planner vs safety) instead of trying to cover everything at once.