## TL;DR
Autonomous driving is shifting from modular pipelines to end-to-end neural approaches. Tesla FSD V12 replaced 300K+ lines of hand-written code with a neural network; Waymo EMMA unifies perception and planning via multimodal models.

## Core Explanation
Traditional modular approach: (1) perception (object detection, tracking, mapping), (2) prediction (intent and trajectory of other agents), (3) planning (path planning, decision making), (4) control (steering, throttle, brake). Each module is independently trained with human-defined interfaces. End-to-end approach: raw sensor data in → driving commands out, with a single neural network learning the entire mapping.

## Detailed Analysis
Imitation learning trains networks on human driving demonstrations (steering angle, speed). Waymo EMMA processes raw camera + user command ("turn right") through Gemini to output trajectory waypoints. Key challenges: causal confusion (correlation ≠ causation), domain shift (training vs deployment distribution), and safety verification of black-box neural controllers.

## Further Reading
- nuScenes and Waymo Open Dataset
- CARLA Simulator
- CVPR 2025 Autonomous Driving Workshop