Unit 14: Computer Vision and Perception

A camera gives the robot measurements arranged as pixels. Everything after that is interpretation.

Follow the evidence from light to action

Pixels Are Measurements starts with image data and coordinates. Detection Is Not Understanding separates model output from robot policy. Test Perception at Its Boundaries investigates where the result stops being dependable.

This Unit is less about building a flashy vision demo and more about refusing to confuse a confident label with ground truth.

Keep the input beside the output

Produce a perception pipeline and a boundary-test matrix.

Preserve the frame or supplied image, the detection/result, relevant score or confidence, the decision rule, and the resulting robot action. Then change one condition at a time: lighting, distance, angle, occlusion, background, or another relevant variable.

A model failure is a system event

Your architecture should define what happens when perception is missing, stale, or uncertain. Slow down, observe again, switch evidence, stop, or hand control to a person are all possible policies.

No live camera is required. Curated images and recorded outputs are enough for the core analysis.