A hands-on tour from the laser pulse to the simulator that trains the whole self-driving stack. Every page works offline on your phone — tap, drag, and explore. （从激光雷达到仿真世界模型）
Each part builds on the last, walking up the stack from the physical sensor to the data engine that everything depends on. Read them in order, or jump to whatever you need.
The sensor itself: light timed to the nanosecond, point clouds, the five hardware types, their parameters, and 2025–26 market pricing. Fire a pulse and watch the round-trip.
↗What consumes the points: the perception pipeline, SLAM (build a map live and watch it drift), sensor fusion, and the point-cloud representations that carry meaning into a model.
↗How defects become silent model biases. Break a clean cloud six ways in the Corruption Lab — sparsity, occlusion, noise, motion, weather, ghosting — and see what each teaches a model in training vs. evaluation.
↗The robustness paradox, resolved: corrupt the inputs, never the labels. Drop a sensor and compare a naïve model (collapses) against a dropout-trained one (degrades gracefully). How data and simulation work together.
↗The destination. Assemble a synthetic cloud stage by stage — raycast, raydrop, intensity, noise — and close the domain gap to a real reference. Why raydrop is the crux, and how you prove a sim actually works.
▸ Offline: every page is fully self-contained — once loaded, no internet needed. Keep the folder together so the inter-page links work.
▸ Interact: sliders, toggles, and tap-to-expand cards throughout. Drag any 3D point cloud to rotate it.
▸ 中文: hard technical terms carry Chinese annotations in purple, with a glossary in Parts 1–2.