Skip to content
Tech Interview Prep
Technical interview guide

Perception & SLAM

How a robot builds a map of an unknown environment while simultaneously figuring out its own location within it.

Read
45 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed
Relevant for
Robotics Engineer

Scope: ORB-SLAM3, Cartographer, GTSAM, Ceres, OpenCV 4.x, ROS 2 Rolling, Kalibr, PCL, Nav2, TUM RGB-D, KITTI, and evo references reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-10An EKF estimates a robot's 2D position [x, y] in metres. After a measurement update, the position covariance block is P = [[0.04, 0.03], [0.03, 0.09]] m². Which statement correctly describes the uncertainty this block encodes?
MCQ-9An indoor ground robot dead-reckons from a 100 Hz IMU while its wheel odometry is faulted, so there is no absolute reference. Over a 10 s run the along-track accelerometer bias is a constant 0.02 m/s^2, and random-walk noise adds far less position error than the bias over that window. The robot starts from a verified zero velocity at t = 0. Roughly how far has the estimated position drifted along that axis when the run ends?
MCQ-19A warehouse robot runs an EKF at 100 Hz. Its wheel odometry reports forward speed with an unbiased Gaussian error of sigma = 0.05 m/s. The team adds an optical-flow sensor that reports speed at 50 Hz with the same sigma, but its scale factor is calibrated 15% high; the filter config still marks it as an independent unbiased source. On a straight run where the true speed is 1.0 m/s for 10 s, what does the filter most likely report?
MCQ-1A warehouse robot has a prebuilt map of the aisles and localizes against it with a particle filter. A second robot is deployed in a newly leased building where no map exists yet. It drives a 40 m loop, and on returning to its start point the particle filter's pose estimate is 1.8 m off from where it began, with the error growing each lap. The team's fix is to add a mapping module that updates landmark positions as the robot moves, so pose and landmark estimates are refined together. Why does this fix address the growing drift, whereas simply running the same particle filter with more particles would not?
MCQ-4A warehouse robot runs EKF-SLAM with a 2D LiDAR. It drives a 30 m loop past four identical shelving units. After the loop the map holds eight shelf landmarks instead of four, and the estimated pose is 1.2 m from where it started. Each shelf is detected with 0.05 m range noise, and the association gate is a fixed 0.3 m Euclidean distance. What most likely explains the duplicated landmarks?
MCQ-11A sliding-window visual-inertial odometry system keeps the 10 most recent keyframes in its factor graph and marginalizes each keyframe as it leaves the window. The keyframe now being marginalized observed 40 landmarks, 25 of which are also observed by keyframes still in the window. What does this marginalization do to the remaining factor graph?
MCQ-12A 2D pose-graph SLAM back-end has 300 poses in a single connected component and 340 relative constraints (odometry plus loop closures). The information matrix H has one eigenvalue of 3×10⁻⁹; the next smallest is 12.5, and all others are larger. Two runs from different initial guesses converge to maps that differ by a 30° rotation and a 4 m translation, with identical final χ² = 12.4. Which diagnosis and fix is correct?
MCQ-13A 1-D pose graph fixes x₀ = 0 and estimates one unknown pose x₁ from three unit-weight edges: two odometry edges measuring x₁ = 1.0 m and x₁ = 1.2 m, and one loop-closure edge measuring x₁ = 5.0 m that is a false positive. The loop-closure edge is given a Huber kernel with δ = 0.5 m, so its cost is r² for |r| ≤ δ and 2δ|r| − δ² otherwise. Where does the optimizer converge for x₁?
MCQ-14A warehouse robot runs pose-graph SLAM down visually identical aisles. Near the end of a run, the loop-closure module matches the current scan to a keyframe from a different aisle, about 20 m back along the trajectory. The match score is high, so the edge enters the graph at full weight even though it disagrees with the odometry chain by several metres. What does that edge do to the map?
MCQ-22A warehouse robot running ORB-SLAM3 revisits an aisle after 40 m. Place recognition matches the current keyframe to keyframe 12 from 90 s ago with appearance similarity 0.91; the runner-up is 0.89. RANSAC PnP over 480 putative matches finds 14 inliers, all confined to a 0.4 m x 0.3 m patch on one pallet label, with 1.2 px mean reprojection error. The recovered relative pose differs from the wheel-odometry prediction by 6.8 m translation and 41 deg yaw. What should the system do with this loop candidate?
MCQ-23A stereo visual SLAM system loses tracking during a fast pan. Relocalization against the map returns a candidate pose from PnP/RANSAC: 150 of 200 matched features are inliers with a median reprojection error of 1.1 px, but all 150 inliers fall inside a 40×40 px patch of repeated floor tiling. The candidate sits 2.8 m and 35° from the last tracked pose. What should the system do before resuming motion?
MCQ-18A monocular SLAM pipeline is evaluated on a 1.2 km test sequence. ATE RMSE with SE(3) alignment (rotation and translation only) is 8.4 m; with Sim(3) alignment (scale also estimated) it drops to 0.35 m. What does this pair of numbers establish?
MCQ-5A warehouse robot runs monocular visual odometry from one calibrated forward camera, with no wheel odometry and no IMU. Scale comes from a single config value: the translation between the first keyframe pair is declared to be 0.5 m, and every feature depth and camera translation is triangulated from it. On a recorded 200-frame run the pipeline reports a path length of 10.0 m. An engineer edits that constant to 0.25 m and re-processes the same recorded images, changing nothing else. What does the pipeline report?
MCQ-6A monocular visual-inertial odometry pipeline tracks FAST corners with KLT optical flow on 640x480 frames at 30 fps. On a warehouse robot it holds 1-2 px reprojection error on textured shelving, but in an aisle of blank shrink-wrapped pallets under a flickering fluorescent ballast the tracked-feature count collapses from ~450 to under 40 within a second, and the surviving tracks drift. Exposure and gain are locked, and the IMU is healthy. What is the most likely cause?
MCQ-8A warehouse robot with a 3D LiDAR drives down a 40 m straight aisle. The aisle has parallel shelving on both sides and a flat floor. Scan-to-map point-to-plane ICP reports a low residual and a high inlier count, but the estimated pose drifts steadily along the aisle axis while lateral position and yaw stay stable. Which explanation fits this behavior?
MCQ-7A mobile robot registers consecutive 16-beam lidar scans in an office corridor with point-to-point ICP: 45 iterations, 2 cm RMS. The team switches to point-to-plane ICP using normals estimated from k = 10 neighbours. Alignment now converges in 6 iterations to 0.8 cm RMS, but over 30 m of travel the pose estimate slides 1.5 m along the corridor axis while lateral and vertical error stay under 2 cm. The corridor has two long parallel walls, a flat floor and a flat ceiling. What explains the slide?
MCQ-20A warehouse AMR's 2D LiDAR spins at 10 Hz and fires 1800 returns sequentially and evenly across each 100 ms revolution; each revolution is written into the point cloud with a single pose. Stamping revolutions with the start-of-sweep pose left a straight wall 5 m away bowed during in-place turns at 90 deg/s. The team now stamps each revolution with the pose the robot held at the middle of the sweep instead. For a return at 5 m range, what is the largest remaining displacement, in metres, between its recorded position and its true position on the wall?
MCQ-3A warehouse AMR carries a 2D lidar and a forward-facing camera. The lidar's scan is matched against the facility map and the robot localizes well, but when the perception stack projects lidar returns into the camera image to fuse them, the projected points land about 40 px to the right of the objects they belong to, and the offset is roughly constant across the whole image. The camera's own calibration was re-run last week and its reprojection error is 0.3 px. What is the most likely cause?
MCQ-2A mobile robot fuses a 10 Hz lidar with a 30 Hz camera. Parked, the estimated lidar→camera extrinsic translation is stable to ±2 mm. Driving straight at 1.5 m/s, the same estimator reports that translation shifted 4.5 cm along the direction of travel; reversing flips the sign, and the magnitude scales linearly with speed. Each sensor timestamps with its own clock, and the lidar's clock runs 30 ms behind the camera's. Features are sharp in every frame. What is the correct diagnosis and fix?
MCQ-17A warehouse AMR localizes by matching its 2-D lidar scan against a map built six months ago by a survey robot. After a repair that re-mounted the lidar, mean lateral error against ground truth rises from 6 cm to 41 cm. The error is near zero at the map origin and grows roughly linearly with distance from it, and rotating every scan by +2.0° in software brings the mean error back to 7 cm. The aisle layout and the map file's checksum are unchanged. What is the most likely cause?
MCQ-15A differential-drive robot runs AMCL against a 2D lidar map. Wheel odometry publishes `odom → base_link` at 50 Hz; AMCL publishes `map → odom` at 10 Hz. At the start of the run `map → odom` is identity. The robot drives 12 m down an aisle and stops. Wheel odometry now reports the robot at (12.4, 0.2) in the `odom` frame, while AMCL's scan match places it at (12.0, 0.0) in the `map` frame. Both transforms are pure translations (no yaw). What does AMCL publish as the new `map → odom` transform?
MCQ-24A warehouse AMR publishes its pose at 50 Hz. At t=142.30 s the published pose jumps +0.82 m in x in a single cycle, then stays consistent. The logs show: wheel odometry continuous with no gaps; IMU at 200 Hz with no dropped samples; a loop-closure factor with residual 0.81 m added to the graph at t=142.28 s; the last relocalization service call at t=98.1 s; and the map->odom transform updated by +0.82 m at t=142.30 s by the same publisher that has owned it all session. Which explanation does the evidence support?
MCQ-21An AMR runs stereo-inertial SLAM at 30 fps, one frame every 33.3 ms, on a single worker thread. Tracking costs 24 ms per frame in the open aisle, but 45 ms for a 60-second stretch where shrink-wrapped pallets fill the view. Every frame goes into an unbounded FIFO and the controller reads the newest published pose, faulting when that pose is more than 80 ms old. About 12 seconds into the stretch the queue holds 94 frames and the pose the controller reads was captured over four seconds earlier. Which change holds the controller's input inside the 80 ms budget for the rest of the stretch?
MCQ-16An AMR localizes with a particle filter fusing wheel odometry, IMU and 2D LiDAR scan matching. Driving down a 40 m aisle lined with identical pallet racks spaced 3 m apart, the scan-match inlier ratio falls from 0.9 to 0.15 and the particle cloud splits into two clusters 3 m apart with weights 0.31 and 0.29, where the previous best cluster held 0.95. The robot is at 1.2 m/s, 6 m from the next cross-aisle, and the aisle half-width is 1.5 m; the stack must keep the robot inside the aisle. What should the navigation stack do?
MCQ-25A visual-inertial SLAM stack for warehouse AMRs is signed off on a 90 s clean corridor run: mean ATE 3.2 cm. In the field, 1 run in 20 loses tracking at the 180° turn into a feature-poor aisle at t ≈ 14 s, and the pose jumps 0.8 m before relocalising. Which change to the validation suite would have caught this?