Name: Autonomous Driving System Architecture: A Comprehensive Analysis of Perception, Decision, Planning and Control
Author: step-3.5-flash

Introduction to Autonomous Driving: Definition, Classification and Development History

In 2024, residents of Phoenix, Arizona, can hail a driverless taxi via a mobile app, with no human safety operator behind the wheel, for a ride to the airport, downtown, or their neighborhood grocery store. In Wuhan, China, more than 1,000 autonomous vehicles operate across the city’s urban roads, core districts, and airport expressways, having completed over 100 million paid rides as of mid-2024. In San Francisco, Waymo’s fully autonomous ride-hailing service now serves 90% of the city’s 49 square miles, with rides priced competitively against traditional ride-sharing platforms. These once-futuristic scenarios are now part of daily life for millions of people, marking a pivotal inflection point in the 120-year history of autonomous driving technology. As a core pillar of the next-generation intelligent transportation system, autonomous driving is not only reshaping how people and goods move, but also redefining urban governance, energy structures, and social equity frameworks. This book aims to provide a comprehensive, structured, and accessible analysis of the core technical architecture of autonomous driving systems, covering perception, decision-making, planning, and control modules, as well as their supporting infrastructure, regulatory frameworks, and commercial applications. As the opening chapter of this book, we will first define the core concept and technical scope of autonomous driving, introduce the globally recognized SAE J3016 classification standard, trace the century-long development trajectory of the technology from conceptual germination to large-scale commercial exploration, analyze its multi-dimensional social value, and outline the core content framework of the full book to help readers establish a holistic cognitive foundation for subsequent in-depth study.

1.1 Definition and Technical Scope of Autonomous Driving

1.1.1 Core Definition of Autonomous Driving

Autonomous driving, also referred to as automated driving or driverless driving, refers to a vehicle system capable of performing all dynamic driving tasks—including environment perception, high-precision positioning, behavior decision-making, motion planning, and vehicle control—without human intervention under specified operating conditions. Unlike traditional human-driven vehicles or advanced driver-assistance systems (ADAS), autonomous driving systems can independently complete the entire closed loop of "perceive-decide-execute" without human input, and bear corresponding legal and safety responsibilities during operation.

The technical boundary between autonomous driving and ADAS is clearly defined by the level of driving automation: ADAS covers L0 to L2 levels, where the system only provides auxiliary functions such as warnings or partial control, and the human driver is always the primary entity responsible for driving safety. In contrast, autonomous driving covers L3 to L5 levels, where the system assumes primary responsibility for driving safety within its designated operational design domain (ODD), which refers to the specific range of operating conditions (including road types, speed limits, weather conditions, and geographic scope) that the system is designed to handle safely[^1].

1.1.2 Core Technical Modules of Autonomous Driving Systems

A complete autonomous driving system consists of five core technical modules, which work together to realize the closed loop of autonomous driving:

1. Environment Perception Module: Often called the "eyes and ears" of the autonomous vehicle, this module collects multi-source environmental information through on-board sensors (including cameras, LiDAR, millimeter-wave radar, and ultrasonic radar) and roadside infrastructure (via V2X communication). It processes the raw data to identify and track surrounding objects (vehicles, pedestrians, traffic signs, traffic lights, etc.), build a 3D environmental model, and provide accurate environmental information for subsequent decision-making. 2.

2. High-Precision Positioning and Simultaneous Localization and Mapping (SLAM) Module: While the perception module identifies surrounding objects, this module determines the vehicle's own precise location within the environment. Unlike standard GPS navigation with meter-level accuracy, autonomous driving requires centimeter-level positioning accuracy to safely navigate lanes, avoid obstacles, and execute precise maneuvers. This is typically achieved through a fusion of Global Navigation Satellite System (GNSS) with Real-Time Kinematic (RTK) correction, Inertial Measurement Unit (IMU) data, wheel odometry, and pre-built high-definition (HD) map matching. For environments where satellite signals are blocked (such as tunnels or underground parking lots), LiDAR-based SLAM or visual SLAM algorithms are used to build real-time local maps and track the vehicle's position without external signals. For example, Waymo’s autonomous vehicles use a combination of RTK-GNSS, high-frequency IMU, and LiDAR SLAM to achieve positioning accuracy of less than 10 centimeters even at highway speeds of 120 km/h, which is critical for safe lane changing and obstacle avoidance at high speeds[^2]. 3. Decision-Making Module: Often referred to as the "brain" of the autonomous vehicle, this module analyzes the environmental and positioning information provided by the previous two modules to make high-level behavioral decisions in line with traffic rules and safety principles. It addresses questions such as "Should the vehicle change lanes?", "Should it yield to a pedestrian crossing the road?", or "Should it stop at a red light?". Unlike rule-based decision systems used in early autonomous driving prototypes, modern decision-making modules increasingly leverage deep reinforcement learning and large language models (LLMs) to handle complex, ambiguous scenarios such as unprotected left turns at busy intersections, yielding to emergency vehicles, or navigating temporary construction zones. For instance, Baidu Apollo’s decision-making module uses a hybrid architecture of rule-based systems and reinforcement learning models, which has reduced the rate of decision-making errors in complex urban scenarios by 37% compared to pure rule-based systems in 2023 testing[^3]. 4. Motion Planning Module: Once the decision-making module determines the vehicle's high-level behavior (e.g., "change to the left lane"), the motion planning module generates a detailed, drivable trajectory that specifies the vehicle's position, speed, and acceleration at every time step over a short-term horizon (typically 3 to 10 seconds). The planning module must balance three core objectives: safety (avoid collisions with all obstacles), comfort (minimize abrupt acceleration, deceleration, and steering maneuvers), and efficiency (reach the destination on time while complying with traffic rules). Motion planning is typically divided into two layers: global planning, which generates a route from the starting point to the destination based on HD map data and real-time traffic information, and local planning, which generates a smooth, collision-free trajectory based on real-time environmental information. For example, when an autonomous vehicle encounters a double-parked delivery truck on its planned route, the local planner will first slow down to maintain a safe following distance, then generate a trajectory to safely change lanes around the truck, and finally merge back into the original lane once the obstruction is cleared. 5. Vehicle Control Module: The final link in the autonomous driving closed loop, this module receives the trajectory generated by the planning module and converts it into concrete control commands for the vehicle's actuators, including the steering system, throttle, and brake. The