Seeing

Every policy in module 04 has a camera in it. This module explains that half of the pipeline: how images become features, and how a network is trained at all.

Do this

• Practical Deep Learning for Coders (fast.ai). Top-down, code-first, free. Lessons 1 to 5 are enough here. It's the most approachable serious course on training networks.
• CS231n: Deep Learning for Computer Vision, course notes (Stanford). The notes, not the lectures, for convolutional networks, training tricks and transfer learning. Read the sections on image classification, optimisation and CNNs.

Build

Train a classifier that tells whether the gripper is holding the block or not, from your module 04 camera frames. Then look at which frames it gets wrong. That's a day's work and it teaches you more about your dataset than any metric.

Skip

• Architecture zoos. One CNN and the idea of a transformer are enough until you need more.
• Building a GPU rig. Rent first.

Done when: you've trained a small vision model on your own images, and you can read a training curve and say what's wrong with it.