Autonomous driving and transportation

Overview

The theoretical foundations of autonomous vehicles (AVs) stretch back to the 1980s with Ernst Dickmanns' VaMoRs project: using saccadic computer vision to steer a Mercedes van: and CMU’s Navlab, which utilized ALVINN, a pioneering neural network for vehicle steering. However, modern AV architecture was catalyzed in the crucible of the DARPA Grand Challenges.

The 2004 Grand Challenge was a deterministic failure. Rule-based edge detection completely collapsed in stochastic environments, confusing tumbleweeds with boulders. The paradigm shifted radically in the 2005 Grand Challenge, where Sebastian Thrun’s "Stanley" won by introducing probabilistic robotics. Discarding monolithic logic, Stanley used a modular, asynchronous message-passing architecture. It relied on an Unscented Kalman Filter (UKF) to fuse odometry, IMU, and GPS data, mapping terrain into a 2D probabilistic occupancy grid. By the 2007 Urban Challenge, CMU’s "Boss" proved AVs could parse complex traffic logic, trajectory generation, and multi-agent tracking, though it highlighted the massive computational payload needed for heuristic models.

Today, the industry is split by the Great Architectural Divergence:

• Waymo (Deterministic Sensor Fusion): Focuses on Level 4 autonomy via strict geofencing. It relies heavily on ground-truth Time-of-Flight (ToF) LiDARs, pre-downloaded High-Definition (HD) maps, and explicit cost-function optimization to guarantee safety within a highly constrained Operational Design Domain (ODD).
• Tesla (Probabilistic End-to-End Vision): Aims for generalized Level 5 autonomy mimicking human biology: two eyes (vision) and a brain (neural net). Tesla’s FSD V12 deprecated radar and ultrasonics, replacing >300,000 lines of explicit C++ heuristics with a monolithic spatial-temporal Transformer.

SAE J3016 Levels of Automation: