Autonomous Driving System Overview and Taxonomy

The journey of autonomous driving from laboratory prototypes to commercial deployment spans nearly a century of iterative innovation, driven by breakthroughs in sensing, computing, and artificial intelligence. The earliest recorded experiment with driverless vehicles dates back to 1925, when Francis Houdina demonstrated a radio-controlled car in New York City, operated by a hidden driver in a trailing vehicle sending electromagnetic signals to control steering, throttle, and braking[^1]. While this was not a truly autonomous system, it laid the conceptual groundwork for eliminating human input from vehicle operation. The field remained largely niche until the 1980s, when Carnegie Mellon University launched the Navlab project, a series of increasingly capable autonomous vehicle prototypes. Navlab 5, tested in 1995, completed a cross-country trip from Pittsburgh to San Diego with 98% of lateral control and 81% of longitudinal control handled by on-board systems, achieving a maximum speed of 105 km/h—a milestone that demonstrated the feasibility of highway autonomous driving[^2].

The 21st century marked a turning point for the industry, catalyzed by the Defense Advanced Research Projects Agency (DARPA) Grand Challenges. The 2004 Grand Challenge, held in the Mojave Desert, offered a $1 million prize for any team that could complete a 240 km off-road course; no team succeeded, with the top finisher traveling only 11 km before stalling. The 2005 challenge saw Stanford University’s Stanley vehicle claim the prize by completing the full course in 6 hours and 54 minutes, using a fusion of vision, LiDAR, and radar sensors to navigate rough terrain. The 2007 DARPA Urban Challenge represented a paradigm shift, requiring vehicles to operate in a mock urban environment obeying traffic rules, navigating intersections, yielding to other vehicles, and parking. Carnegie Mellon’s Boss vehicle won the competition, demonstrating that autonomous driving could function in complex, structured road scenarios[^3]. These challenges attracted massive investment from tech companies and automakers, accelerating the transition from research to commercial deployment.

The first wave of commercial autonomous driving products focused on driver-assistance systems (ADAS) under SAE Level 2, with Tesla’s 2015 launch of Autopilot marking a watershed moment for mass-market adoption. By 2023, over 30 million vehicles globally were equipped with Level 2 ADAS features including adaptive cruise control, lane-keep assist, and automated emergency braking. The first commercial Level 3 deployments arrived in 2021, when Mercedes-Benz launched its Drive Pilot system in Germany, allowing drivers to hand over full control of the vehicle on controlled highway segments at speeds up to 60 km/h, with the automaker assuming liability for accidents during system operation[^4]. As of 2024, Level 3 systems have been approved for use in Germany, China, Japan, and several U.S. states, with cumulative global deployment exceeding 1.2 million units. Level 4 commercial robotaxi services, which operate without human drivers in designated operational domains, have been launched in over 20 cities globally, led by Waymo in the U.S., Baidu Apollo in China, and Pony.ai in both markets.

> 💡 电子版侧边栏：Key Definition - Driver Assistance vs. Automated Driving > The line between ADAS (Level 1/2) and automated driving (Level 3+) hinges on responsibility for vehicle operation. In Level 2 systems, the human driver is legally and operationally responsible for monitoring the driving environment at all times, even when the system is active. In Level 3 and above, the automated system assumes operational responsibility within its defined Operational Design Domain (ODD), and the human driver is only required to take over when the system issues a takeover request.

1.2 SAE J3016 Autonomy Levels and Operational Design Domain (ODD)

The SAE International J3016 standard, first published in 2014 and updated most recently in 2021, defines six levels of driving automation, ranging from no automation (Level 0) to full automation (Level 5), with clear boundaries for system capability, driver responsibility, and ODD requirements[^5]. The ODD is a critical concept for all levels above L2: it defines the specific set of conditions under which an automated driving system is designed to operate safely, including geographic boundaries, road types, speed ranges, weather conditions, and time of day. For example, a Level 3 highway pilot system may have an ODD limited to controlled-access highways with clear lane markings, speeds between 60 and 120 km/h, and light rain conditions; outside of this ODD, the system will alert the driver to take control.

Each SAE level has distinct operational and liability characteristics, summarized in Table 1-1. Level 0 (No Driving Automation) covers all conventional human-driven vehicles, with no automated functionality for steering, throttle, or braking. Level 1 (Driver Assistance) provides support for either longitudinal control (e.g., adaptive cruise control) or lateral control (e.g., lane-keep assist), but not both simultaneously, with the driver fully responsible for all driving tasks. Level 2 (Partial Driving Automation) combines longitudinal and lateral control in the same driving scenario, but requires the driver to remain fully attentive and ready to take over at any time; examples include Tesla Autopilot, GM Super Cruise, and NIO Pilot. Level 2 systems are often marketed as "L2+" or "semi-autonomous," but they do not meet the SAE definition of Level 3, as the driver remains legally responsible for all driving outcomes.

Level 3 (Conditional Driving Automation) is the first level where the automated system assumes operational responsibility within its ODD. The driver is not required to monitor the driving environment when the system is active, and may engage in non-driving tasks such as reading or using a phone; however, the driver must be ready to take over within a specified response time (typically 10-30 seconds) when the system issues a takeover request. The 2021 SAE update explicitly clarified that for Level 3 systems, the vehicle manufacturer or system provider is liable for accidents caused by system failure within the ODD, a key enabler for commercial deployment. Level 4 (High Driving Automation) eliminates the need for a human driver entirely within its ODD: even if the driver does not respond to a takeover request, the system will pull over safely or continue operating without human intervention. Level 4 systems are currently deployed in geofenced areas such as urban districts, airports, and industrial parks, with ODDs limited to mapped roads and predefined operating conditions. Level 5 (Full Driving Automation) has no ODD restrictions: the system can operate safely in all road types, weather conditions, and geographies, equivalent to a human driver with unlimited experience. No Level 5 systems have been deployed commercially as of 2024, as the technical challenges of covering all possible driving scenarios remain unsolved.

| SAE Level | System Capability | Driver Responsibility | ODD Requirement | Liability | |----------|------------------|----------------------|----------------|-----------| | L0 | No automation | Full responsibility | None | Driver | | L1 | Single function control (longitudinal or lateral) | Full responsibility | Any | Driver | | L2 | Combined longitudinal and lateral control | Must monitor environment at all times | Any | Driver | | L3 | Full control within ODD | Required to take over on system request | Defined (e.g., highway, daylight) | Manufacturer for system-related incidents | | L4 | Full control within ODD, no takeover required | No responsibility within ODD | Geofenced, predefined conditions | Manufacturer for all incidents within ODD | | L5 | Full control in all conditions | No responsibility | None | Manufacturer for all incidents |

A common misconception in