Module Check: Modeling the World
You have traced the full arc from raw sensor signal to a structured world model. Pixels became object detections. Laser pulses became 3D point clouds. Noisy measurements became filtered pose estimates. Independent sensor streams became fused, reliable data. Observations became persistent maps and semantic representations. And you have seen how uncertainty — irreducible but manageable — threads through every layer. This module check tests whether you can reason with these ideas, not merely recall them. Take your time.
Flashcards — click each card to reveal the answer
Module Quizzes
A robot is given a 2D lidar scan and a known map of the environment. It must estimate its pose. A Kalman filter initialized at (x=0, y=0, theta=0) is used, but the robot was actually placed at (x=5, y=3, theta=90 degrees). What will happen?
A stereo camera pair is used for depth estimation on a robot navigating a white-painted hallway. The depth estimate in the center of the hallway is unreliable while depth near the baseboards and ceiling is accurate. The most likely cause is:
A particle filter for robot localization starts with 1000 particles uniformly distributed across a large building floor plan. After 30 seconds of motion and lidar observation, only 3 particles remain in a small cluster near the correct location. The filter then resamples to 1000 particles. Which statement is most accurate?
An autonomous vehicle fuses lidar point clouds with camera images to produce labeled 3D detections. The lidar runs at 10 Hz and the camera runs at 30 Hz. The system uses the most recent lidar scan and the most recent camera image for fusion. At 60 mph (26.8 m/s), what is the maximum position error introduced by temporal misalignment between a lidar scan and a camera frame that are up to 33 ms apart?
A robot's semantic map records a fire exit door as 'closed' based on a camera observation made 8 minutes ago. During an emergency evacuation, the robot routes other robots away from this exit. What is the fundamental perception design failure here?
A designer proposes to replace a robot's 64-beam lidar with a stereo camera to reduce cost and weight. The robot operates in an outdoor environment, navigating at 1 m/s. Which tradeoff analysis is most technically complete?
Capstone Synthesis
Capstone: Audit a Real Robotic Perception Failure
- This capstone asks you to analyze a real-world robotic perception failure using the full conceptual framework of this module. Choose one of the following publicly documented incidents, or propose an equivalent with instructor approval:
- (A) A 2016 Tesla Autopilot fatality in which the system failed to detect a white semi-truck trailer against a bright sky (camera-based perception).
- (B) A 2018 Uber ATG fatality in which a lidar-camera fusion system failed to correctly classify a pedestrian pushing a bicycle outside a crosswalk.
- (C) A Mars rover (Spirit or Opportunity) that became stuck due to terrain assessment failure — the visual terrain model indicated safe ground where the soil was unexpectedly soft.
- Your analysis must address all of the following:
- 1. Perception pipeline failure: At which layer of the perception stack did the failure originate? Was it sensing, perception algorithm, state estimation, or world model? Provide a technical account of what went wrong at that layer.
- 2. Uncertainty handling: What information was available to the system that should have indicated uncertainty? Was that uncertainty propagated to the planner? If not, why might the system have presented high confidence instead?
- 3. Sensor fusion diagnosis: Could an additional or different sensor have detected what was missed? Be specific — name the sensor type, explain what measurement it would have made, and describe how fusion with the primary sensor's output would have changed the perception result.
- 4. World model diagnosis: Did the world model contain a representation of this object or situation? If the representation was absent or wrong, what design decision caused the gap?
- 5. Systemic lessons: Beyond this specific incident, what general design principle does this failure illustrate? State it as a rule that could be applied to future perception system design.
- 6. Ethical dimension: Autonomous systems sometimes fail in ways that kill people. Given the analysis above, do you believe the system should have been deployed at the level of autonomy it was operating at when the failure occurred? Support your answer with the technical evidence you have analyzed.
- Write your analysis as a structured technical-ethical report. Be rigorous. Be honest about uncertainty in your own analysis where the public record is incomplete.