Robot perception

1.In a pinhole camera model, what do the intrinsic parameters describe?

2.A camera has fx = 800 pixels and cx = 640 pixels. A point sits in the camera frame at X = 0.5 m (right) and Z = 4 m (forward). Ignoring distortion, what is its horizontal pixel coordinate u?

3.A team halves the resolution of a calibrated camera stream from 1280x960 to 640x480 to save compute. How should they update the intrinsics?

4.A hospital delivery robot's camera calibration reports 0.2 pixel reprojection error. Yet in rectified images taken while the robot is parked, straight door frames still bend near the image corners. What is the most likely cause?

5.The yaw in a lidar-to-camera extrinsic is off by 0.5 degrees. About how far sideways will a lidar point at 40 m land from the object it belongs to?

6.After a sidewalk delivery robot bumps a curb, lidar points projected onto its camera image sit shifted left of every object, by about the same pixel amount across the frame. What do you do first?

7.A quadruped inspecting a pipeline carries a 190-degree lens. Which camera model should calibration use?

8.An orchard drone yaws quickly while filming, and the vertical tree trunks appear slanted in each frame. The trunks look straight when it hovers. What is the most likely cause?

9.A camera image is 1280 pixels wide and its horizontal focal length is fx = 640 pixels. With the principal point at the center, what is the horizontal field of view?

10.A runway debris robot uses a camera with fx = 1000 pixels. About how many pixels wide is a 0.3 m bolt bag lying 30 m away?

11.A lab pipetting arm carries a camera on its wrist. Which calibration finds the fixed transform between the camera and the arm's flange?

12.A rectified stereo pair has focal length 700 pixels and a 0.12 m baseline. A point shows a disparity of 21 pixels. What is its depth?

13.For a stereo camera with a fixed disparity error of a fraction of a pixel, how does depth error change with distance?

14.A stereo rig has f * B = 84 pixel-meters and a disparity error of 0.25 pixels. About how large is the depth error for a point at 8 m?

15.A warehouse robot's stereo camera gives depth that is too noisy for pallets at 15 m. Which single hardware change helps most?

16.A hotel room-service robot uses passive stereo. Depth is missing or wildly wrong on long, plain painted walls. Which addition helps most?

17.What does stereo rectification guarantee for the matching step?

18.A greenhouse picking robot uses stereo depth. Rows of evenly spaced trellis wires sometimes show up as phantom obstacles much closer than the wires really are. What is the most likely cause?

19.What is the purpose of a left-right consistency check in stereo matching?

20.What does voxel grid downsampling do to a lidar point cloud?

21.A construction site rover removes ground with one RANSAC (random sample consensus) plane fit per scan. On hilly ground, the slope ahead keeps getting flagged as an obstacle. What is the best fix?

22.A Euclidean clusterer with a fixed 0.3 m distance tolerance works well near the robot, but far pedestrians break into several small clusters. What is the most likely cause?

23.A lidar has 0.2 degree horizontal resolution. About how many returns per laser ring land across a 0.5 m wide person standing 50 m away?

24.A 32-beam lidar spreads its beams evenly across a 40 degree vertical field of view. About how far apart vertically are neighboring beams at 50 m?

25.At a parcel hub, a sorting arm clusters lidar points to find boxes. Two boxes touching on the belt keep coming out as one cluster. What is the best approach?

26.A night delivery van's lidar detects a black tire on the road much later than a white box the same size. What is the most likely reason?

27.On a mining haul truck, the lidar reports a diffuse cloud of low-intensity points in front of the truck. It drifts and vanishes within a few frames. What is it most likely to be?

28.A lawn mower robot downsamples its lidar with 0.5 m voxels before obstacle detection. It keeps running over fallen branches. What change do you make?

29.What does an automotive-style radar measure directly that a camera and a lidar do not?

30.A campus shuttle's radar misses a cyclist crossing straight across its path, but it tracks cyclists riding ahead of it fine. What is the most likely cause?

31.An FMCW (frequency-modulated continuous wave) radar sweeps 1 GHz of bandwidth. What is its range resolution?

32.A highway truck's radar reports a single target where two cars drive side by side 100 m ahead. Why can't it separate them?

33.A straddle carrier at a container port drives between stacks of steel containers. Its radar shows targets that appear to be inside or behind the stacks. What is the most likely cause?

34.Why do many radar pipelines have trouble reporting a stopped car ahead?

35.A snowplow drives through a blizzard at night. Its camera is blinded and its lidar fills with snowflake returns. Which sensor is most likely to still detect a car ahead?

36.A vehicle drives at 20 m/s. Its camera timestamps run 50 ms behind its lidar timestamps. About how far apart will matching detections land when fused?

37.A delivery van's spinning lidar takes 0.1 s per sweep while the van drives at 15 m/s. If points are not motion-compensated, how far does the van move between the start and the end of one sweep?

38.A robot's camera driver stamps each frame with the time it arrives at the host. Why is this a problem for fusion?

39.A robot turns at 1 rad/s. Its radar and camera clocks are offset by 30 ms. About how far sideways does a target 20 m away shift between the two sensors' views?

40.A global shutter camera uses a 10 ms exposure. Which time should a fusion pipeline assign to the frame?

41.On a pipeline inspection quadruped, fused camera and lidar boxes line up when the robot stands still. They drift apart whenever it turns. What is the most likely cause?

42.Why do some teams trigger a camera from the angle of a spinning lidar?

43.Two sensors use free-running clocks that drift apart by 50 parts per million. About how far apart are they after one hour without resync?

44.What separates early fusion from late fusion in multi-sensor perception?

45.Which is a real advantage of late fusion over early fusion?

46.A forklift in a cold storage warehouse needs to spot pallets that give weak lidar returns and low-contrast images. Why might early fusion help here?

47.A late fusion system pairs camera detections with lidar clusters. Matching often fails beyond 30 m, even though both sensors see the object. Which matching method is most robust?

48.Two position estimates both come from the same lidar scan through different pipelines. A filter fuses them as if they were independent. What goes wrong?

49.Two independent sensors estimate an object's position as 10 m (variance 1) and 15 m (variance 4). What is the minimum-variance fused estimate?

50.A fused camera-lidar detector was trained with both sensors always working. In the field, the camera fails at night and detection collapses, even though the lidar is fine. What do you change in training?

51.A tracker fuses lidar and radar for vehicles on a farm road. What does the radar add that is hardest to get from lidar alone?

52.Two 10 by 10 pixel boxes overlap. One is shifted 5 pixels sideways from the other, with no vertical shift. What is their IoU (intersection over union)?

53.A detector produces 80 true positives and 20 false positives, and misses 40 objects. What are its precision and recall?

54.What does mAP (mean average precision) average in a typical object detection benchmark?

55.A sidewalk robot's detector has high mAP (mean average precision) overall. But recall on people using wheelchairs is poor, and that class has few test examples. Why did mAP hide this?

56.A parking garage robot is evaluated with an IoU (intersection over union) threshold of 0.5. Small, far objects score badly even though the boxes look right by eye. Why?

57.You raise a detector's confidence threshold from 0.3 to 0.7. What usually happens?

58.A shelf-scanning robot's detector scores well on per-frame mAP (mean average precision). On the robot, the planner keeps swerving because a nearby cart is found in some frames and missed in others. What is missing from the evaluation?

59.In 3D detection, which error does 3D IoU (intersection over union) fail to penalize?

60.A product team says phantom braking from false detections is their top complaint. Which metric should you add to the detector report?

61.A rail yard inspection robot's detector shows 95 percent recall overall. Split by range, recall is 99 percent within 20 m and 70 percent from 30 to 50 m. What should you conclude?

62.In multi-object tracking, what does the Hungarian algorithm give you over greedy nearest-neighbor matching?

63.Why do trackers gate associations with Mahalanobis distance instead of plain Euclidean distance?

64.A mall security patrol robot tracks shoppers. When two people cross paths, their track IDs often swap. Which change helps most?

65.A tracker creates a new track for every unmatched detection. False tracks keep popping up from single-frame noise. What is the usual fix?

66.An airport baggage tug tracks people walking behind parked carts. A person hidden for about one second comes back with a new track ID. What change helps most?

67.Which tracking metric focuses most on whether each object keeps the correct identity over time?

68.A robot tracks objects in its own body frame. When it turns in place, parked cars get tracks with large velocities. What is wrong?

69.Tracks on forklifts overshoot whenever the forklifts turn sharply, then slowly catch up. The tracker uses a constant velocity model. What helps most?

70.A late detection arrives, stamped 80 ms earlier than the track's latest update. What is the correct way to handle it?

71.A 1D Kalman filter predicts a position of 10 m with variance 4. A measurement reads 12 m with variance 4. What are the updated mean and variance?

72.A 1D Kalman filter has prior variance 1 and measurement variance 9. What is the Kalman gain?

73.What does the state covariance in a Kalman filter represent?

74.A tracked object goes unseen for several frames, so the filter runs only prediction steps. What happens to the covariance?

75.A tracker for a vineyard robot follows workers well when they walk steadily. When a worker stops suddenly, the track keeps sliding forward for a second. Which tuning is most likely wrong?

76.A filtered position for a pool cleaning robot jitters almost as much as the raw sensor readings. Which tuning is most likely wrong?

77.How can you check whether a tracking filter's covariance is honest on real data?

78.A tracker keeps a Cartesian state but receives radar range and bearing measurements. Why use an extended or unscented Kalman filter?

79.A beach cleaning robot faces the low evening sun. Its camera saturates and the detector misses a person ahead. What is the best design response?

80.In heavy rain, what is the most common effect on a lidar?

81.An office cleaning robot keeps bumping into a glass partition. Its lidar costmap shows free space where the glass is. What is the most likely cause?

82.A street robot's lidar shows a halo of phantom points around a highway sign. What is the most likely cause?

83.A security patrol robot works a parking lot at night. Its camera uses long exposures to see, and walking people become smeared streaks. What is the most likely tradeoff here?

84.After rain, a campus robot's camera detector reports upside-down people below the ground. What is the most likely cause?

85.A ski resort snow groomer's front lidar slowly loses returns over a shift. The points still present look clean. What should the perception system have?

86.A search and rescue drone uses a long-wave thermal camera to find people at night. Which limit should the team plan for?

87.Why does visual odometry drift over time?

88.A visual odometry system drifts about 1 percent of distance traveled. About how far off could its position be after a 500 m route with no loop closure?

89.A warehouse mapping robot's SLAM (simultaneous localization and mapping) map suddenly folds two aisles on top of each other. What is the most likely cause?

90.Why can't monocular visual odometry recover metric scale on its own?

91.A visual odometry system on a train platform robot drifts badly at rush hour but works well when the platform is empty. What is the most likely cause?

92.A lawn mower robot uses wheel odometry to localize. On wet grass, its map position drifts far ahead of where it really is. What is the most likely cause?

93.A vineyard robot localizes against a lidar map built in winter. In summer, with full leaves, localization keeps failing. What should the team do?

94.Why do many visual odometry systems add an IMU (inertial measurement unit)?

95.A monocular visual odometry system fails when a pan-tilt camera rotates in place without moving. What is the reason?

96.Two labeling teams handle the same dataset. One labels partly hidden pedestrians with the full body box, and the other boxes only the visible part. What do you do first?

97.An orchard harvester's fruit detector was trained only on sunny summer data. The team will deploy it in autumn rain. What should they do before deploying?

98.A team randomly splits frames from their driving logs into train and test sets. Test metrics look excellent but drop sharply in new areas. What is the most likely issue?

99.Human labels for 3D box centers vary by about 10 cm between labelers. A team reports a model that improved center error from 9 cm to 8 cm. How should you read this?

100.A team has a large stream of unlabeled field data and a small labeling budget. Which frames should they send for labeling first?

101.A team auto-labels new data with a large offline model and trains the onboard model on those labels. What is the main risk?

102.A rover meets an object type that was never in its training data, such as an overturned shopping cart. How should the perception stack handle it?

103.A team tests a rare scenario 30 times and sees zero failures. What can they say, at about 95 percent confidence, about the true failure rate?

104.A delivery drone team wants to know how its detector does on rare cases like kites and balloons. Their test set has only a few examples of each. What do you do?

105.A robot drives at 2 m/s. Its perception pipeline takes 150 ms from capture to output. How far does the robot move before it can react to what was seen?

106.A camera runs at 30 frames per second, so an event can wait up to 33 ms before the next capture. Transfer takes 10 ms, inference 45 ms, and tracking 5 ms. What is the worst-case age of the event at the tracker's output?

107.A perception pipeline averages 60 ms of latency, but the 99th percentile is 250 ms. Which number should drive safety analysis?

108.Tracked objects reach the planner 120 ms after their data was captured. How should the system use them?

109.To raise GPU throughput, a team batches 4 camera frames per inference call at 20 fps. What happens to perception latency?

110.A sorting arm at a recycling plant picks items off a belt moving at 1.5 m/s. Detection takes 80 ms and the arm needs 120 ms more to start its grasp. Where should the arm aim?