One page per project. Each explains what the system does in plain language, walks through every source file, sets out the numbers that were measured and how, and states what the result does not prove. No prior knowledge of the specific tools is assumed — each page opens with a glossary of its own terms.
An anchor-free detector finds objects on a table where they hide one another. COCO mAP is implemented from scratch rather than imported, and an ablation compares photometric augmentation against randomising appearance at render time — the two are not symmetric.
The same cart-pole solved twice: once with an optimal linear controller derived from the physics, once with PPO written out from first principles. Sixteen seeds, four ablations and a permutation test decide where the learned policy is genuinely better and where it is not.
A walking policy for a small biped, trained without a GPU. The project opens by measuring whether that is possible at all — four wrong throughput numbers preceded the right one — then fixes a 48-dimensional observation contract and a hand-tuned controller for learning to beat.
A small CNN recovers a cube's position and yaw angle from one rendered image. Two classical baselines come first — the second exposed a systematic radial bias in the first, fixed by a single scale factor — so the network's error is measured against something that already works.
A 7-axis Franka Panda clears seven balls off a table. Nothing is hard-coded: a simulated camera renders the scene, an OpenCV node finds the balls in the picture, a C++ node discards anything the arm cannot reach, and a state machine grasps only what survived that chain.
Three simulated robots publish where they are. A message bus carries it to two independent readers — one stores every reading, one notices when a robot has stopped and opens an alert a human can close. The whole fleet is described in one configuration file, and a job on every push brings the real container stack up and proves the chain end to end.