Overview of Autonomous Driving Systems: Classification, Architecture and Development Milestones

The global road transportation system faces a profound and persistent crisis: according to the World Health Organization’s 2023 Global Status Report on Road Safety, approximately 1.35 million people die annually in traffic collisions, with 94% of these crashes attributable to human error, including distracted driving, drunk driving, and fatigue[^2]. Autonomous driving systems (ADS), designed to assume partial or full responsibility for the dynamic driving task (DDT) — the real-time operational functions of vehicle control, environmental monitoring, and tactical response required for on-road operation — hold the transformative potential to reduce crash rates by orders of magnitude, reshape urban mobility ecosystems, and unlock new economic value across logistics, passenger transport, and related industries. As a foundational text for engineers, researchers, and industry practitioners seeking to master the core technical stack of ADS, this chapter first formalizes the universally accepted classification framework for autonomous driving, then deconstructs the canonical layered architecture of production-grade systems, and finally traces the technology’s evolution from experimental research prototypes to commercial deployment, while outlining the core bottlenecks and development trajectories that define the current stage of the industry.

SAE J3016: Canonical Classification Framework for Autonomous Driving #SAE-J3016 #Classification #Autonomous-Driving-Levels

The global industry relies on the SAE J3016 standard, published by SAE International (formerly the Society of Automotive Engineers), as the authoritative reference for classifying autonomous driving capabilities[^1]. First released in 2014 and most recently updated in April 2021, the standard defines five distinct levels (L0 to L5) based on two core, unambiguous dimensions: (1) the entity responsible for executing all subtasks of the DDT, including environmental perception, path planning, vehicle control, and system monitoring; and (2) the entity responsible for performing fallback maneuvers when the system reaches its operational limit, experiences a fault, or is requested to cede control. This dual-dimension framework eliminates the ambiguity that plagued earlier marketing-driven classification schemes, and clearly demarcates the boundary between driver assistance systems, which require constant human oversight, and fully autonomous systems, which assume operational responsibility within defined constraints.

Level 0: No Driving Automation

At L0, the human driver bears full responsibility for all DDT subtasks. The vehicle may provide passive alerts (e.g., lane departure warnings, forward collision warnings) but exerts no active control over longitudinal or lateral vehicle motion. All current production vehicles without advanced driver assistance systems (ADAS) fall into this category, as do vehicles equipped only with basic safety alert features. For example, a 2022 Toyota Corolla without adaptive cruise control (ACC) or lane-keep assist (LKA) is classified as L0, even if it is equipped with electronic stability control (ESC), which is a passive safety feature rather than an automation feature.

Level 1: Driver Assistance

L1 systems take over a single, discrete DDT subtask, either longitudinal (acceleration/braking) or lateral (steering), while the human driver retains full responsibility for all other subtasks and for monitoring the driving environment at all times. The most ubiquitous L1 feature is ACC, which automatically adjusts vehicle speed to maintain a set following distance from the preceding vehicle, but requires the driver to control steering and monitor for obstacles. Many mid-tier 2024 production vehicles combine ACC and LKA as a bundled L1 feature set, but the driver remains legally and operationally responsible for all driving tasks, and must be prepared to intervene at any time.

Level 2: Partial Driving Automation

L2 systems simultaneously control both longitudinal and lateral vehicle motion under specific conditions, enabling hands-off (but not eyes-off) operation for short periods. Unlike L1 systems, L2 systems can handle both acceleration/braking and steering in parallel, for example maintaining lane position while following a leading vehicle at highway speeds. However, the human driver remains fully responsible for monitoring the driving environment and intervening when the system reaches its operational limit, and must be prepared to take control at a moment’s notice. The vast majority of ADAS systems marketed as “Autopilot,” “Navigate on Autopilot,” or “NGP” by global OEMs are classified as L2 under SAE J3016, including Tesla Autopilot, Xpilot (Xpeng), and AD Max (Li Auto). It is a critical industry misconception that marketing terms such as “L2

+Plus” or “nearly L3” used by many original equipment manufacturers (OEMs) denote a capability beyond L2 per SAE J3016. These labels almost always refer to L2 systems with expanded conditional operation (e.g., urban road lane centering, automated lane changes) but retain the core L2 requirement that the human driver must monitor the driving environment at all times and assume full operational responsibility. For instance, Tesla’s “Full Self-Driving (FSD) Beta” is explicitly classified as L2 by both SAE International and the U.S. National Highway Traffic Safety Administration (NHTSA), despite marketing materials that have historically implied full autonomous capability; in 2022, NHTSA forced Tesla to remove references to “Autopilot” implying full autonomy from its marketing materials following a series of crashes where drivers over-trusted the system[^3]. The key technical and legal distinction between L2 and L3 is not the breadth of operational capability, but the transfer of operational responsibility: in L2, the human driver is always the responsible entity, while in L3, the ADS assumes operational responsibility within its defined operational design domain (ODD), and the human is only required to take over when the system issues a clear takeover request with sufficient lead time (typically 10–15 seconds for highway scenarios).

Level 3: Conditional Driving Automation

L3 marks the first formal transfer of operational responsibility for the DDT from the human driver to the ADS, representing a fundamental shift in the regulatory and safety framework for on-road vehicles. Under L3, the ADS executes all DDT subtasks (perception, planning, control, monitoring) within a strictly defined ODD, and the human driver is not required to monitor the driving environment during automated operation. When the system reaches its operational limit (e.g., approaching a construction zone, encountering heavy rain that exceeds its sensing capability) or experiences a system fault, it must issue a clear, multimodal takeover request (visual, auditory, and haptic alerts) with a minimum lead time of 10 seconds, per the UN Regulation No. 157 (UN R157) for L3 automated driving systems, which has been adopted by the European Union, Japan, South Korea, and other major automotive markets. If the human driver fails to take over within the requested time, the ADS must execute a predefined Minimum Risk Maneuver (MRM): a safe, predictable sequence of actions to bring the vehicle to a complete stop in a low-risk location (e.g., the shoulder of the road) while activating hazard lights.

The first mass-produced L3 system to receive regulatory approval was Mercedes-Benz’s Drive Pilot, launched in 2022 for the S-Class and EQS models in Germany, and later approved for use in the U.S. state of Nevada in 2023. Drive Pilot’s ODD is strictly limited to traffic-jam scenarios on highways or controlled-access roads with physical barriers separating opposing traffic, at speeds below 60 km/h, in clear weather with dry pavement, and with no active construction or pedestrians in the roadway. Within this ODD, drivers are legally permitted to engage in non-drivi