Lesson 1: A vision transformer that reads camera test cards

Module 13 · Lesson 1 · About 50 minutes · Needs: Module 4, Lesson 4 (patches and grid positions); Module 3 (the block); Module 6, Lesson 1 (the training recipe)

The situation

Sol 447 at Aurora Station. Isla Duarte, the robotics engineer, brings a stack of camera test cards: small 32 by 32 pictures of one to three squares, circles or triangles. "Before a robot reaches for anything, its camera has to say what is on the card and how many. Can your lab do that without Earth?"

This lesson builds a vision transformer (ViT) in full, trains it on shape, then on count, and measures what patch size costs. The running example is card 3, "three red circles". Module 4, Lesson 4 built just enough of a ViT to test positions; this is the whole model.