Introduction to Autonomous Vehicle Systems: Evolution, Definitions, and Societal Impact

The conceptual roots of autonomous vehicles (AVs) extend far beyond the digital age, tracing back to the earliest days of motorized transport in the 1920s, when *Popular Mechanics* magazine first published speculative illustrations of “driverless cars” guided by embedded radio signals in roadways. These early visions, while technologically infeasible at the time, captured a fundamental human desire to eliminate the burden of driving, reduce traffic accidents, and expand mobility access to populations unable to operate traditional vehicles. For nearly half a century, these ideas remained in the realm of science fiction, until advances in computing, sensor technology, and artificial intelligence in the late 20th century made the first functional prototypes possible.

The first experimental autonomous vehicle was built by General Motors in 1958, in partnership with researchers at the California Institute of Technology. Dubbed the “Firebird I,” this prototype used a series of magnetic spikes embedded in the road surface to guide its steering, with a radio control system sending feedback to a central console that adjusted the vehicle’s path. While the system only worked on specially equipped test tracks and required constant human oversight, it proved that vehicle guidance without direct human input was physically possible. Over the next two decades, researchers at institutions including Stanford University and the University of Tokyo built on this work, developing early computer vision systems that could detect road lane markings and basic obstacles, but these systems were limited by the bulky, low-power computing hardware of the era, which could only process a few frames of video per second.

A major breakthrough came in 1986, when the Robotics Institute at Carnegie Mellon University (CMU) launched the Navlab (Navigation Laboratory) project, the first on-road autonomous vehicle capable of navigating public roads without dedicated infrastructure. The initial Navlab 1 prototype used a black-and-white CCD camera with a resolution of just 320x240 pixels, paired with a custom-built image processing system that could detect road lane markings at speeds of up to 3 km/h. Over the next decade, the Navlab team iterated rapidly on the platform: Navlab 5, completed in 1995, became the first vehicle to drive across the United States autonomously, completing a 4,600 km journey from Pittsburgh, Pennsylvania to San Diego, California with 98% of the route driven without human input, at an average speed of 105 km/h. For context, the on-board computer used for Navlab 5 had a processing power of just 40 MIPS (million instructions per second), roughly equivalent to a low-end calculator today; modern autonomous vehicle compute platforms operate at thousands of TOPS (trillions of operations per second), a performance improvement of more than 10 million times in less than 30 years.

The field entered a new era of rapid development in the early 2000s, driven by a series of high-profile challenges hosted by the U.S. Defense Advanced Research Projects Agency (DARPA). The 2004 DARPA Grand Challenge offered a $1 million prize to any team that could build an autonomous vehicle capable of navigating a 240 km off-road course in the Mojave Desert. No team completed the course that year, with the most successful entry traveling just 11 km before getting stuck on a rock. The 2005 challenge, however, saw a dramatic improvement: Stanford University’s “Stanley” vehicle completed the full course in 6 hours and 53 minutes, averaging 30 km/h over desert terrain riddled with obstacles, narrow passes, and steep inclines. Stanley used five SICK laser rangefinders, two forward-facing cameras, and a GPS/IMU (Inertial Measurement Unit) navigation system, paired with an on-board Pentium M processor running at 1.6 GHz. The success of the 2005 challenge proved that autonomous vehicles could operate reliably in unstructured, unpredictable environments, and sparked a wave of investment from both tech companies and traditional automakers.

By the early 2010s, Google (now Alphabet) had launched its secretive X Lab autonomous vehicle project, which quickly became the global leader in on-road testing. By 2018, Waymo, the spin-off company focused on commercializing autonomous technology, launched the world’s first fully driverless ride-hailing service in Phoenix, Arizona, operating a fleet of Jaguar I-PACE vehicles retrofitted with custom sensor suites, with no human safety driver behind the wheel. As of 2024, Waymo operates fully driverless services in Phoenix, San Francisco, Los Angeles, and Austin, with more than 100,000 paid rides per week. In China, companies including Baidu Apollo, WeRide, and Pony.ai have launched commercial robotaxi services in more than 20 cities, including Beijing, Shanghai, Guangzhou, and Shenzhen, with more than 5 million cumulative test miles driven on public roads as of 2024.

1.1 Standardized Classification of Automated Driving: The SAE J3016 Framework

As the field evolved from research labs to commercial deployment, a critical need emerged for a standardized, universally accepted definition of what constitutes an “autonomous” vehicle. Without clear definitions, automakers and tech companies could use misleading marketing language to overstate the capabilities of their systems, creating public confusion and safety risks. To address this gap, the Society of Automotive Engineers (SAE) International published the J3016 standard in 2014, which has since become the global benchmark for classifying automated driving systems (ADS).

The 2019 revision of SAE J3016, the most widely adopted version as of 2024, defines six distinct levels of driving automation, ranging from no automation (Level 0) to full automation (Level 5), as illustrated in the diagram below. The core distinction between levels is the division of responsibility for dynamic driving tasks (steering, acceleration, braking, and monitoring the driving environment) between the human driver and the automated system.

```mermaid graph TD L0[L0: No Automation<br/>人类承担所有驾驶任务] --> L1[L1: Driver Assistance<br/>单一功能辅助（如自适应巡航、车道保持）<br/>人类全程监控驾驶环境] L1 --> L2[L2: Partial Driving Automation<br/>多功能组合辅助（如同时跟车+车道保持）<br/>人类必须全程监控并随时接管] L2 --> L3[L3: Conditional Driving Automation<br/>系统在特定场景下承担所有驾驶任务<br/>人类需在系统请求时接管，无需全程监控] L3 --> L4[L4: High Driving Automation<br/>系统在特定设计域（ODD）内全自动驾驶<br/>无需人类接管，ODD外需人类操作] L4 --> L5[L5: Full Driving Automation<br/>全场景全自动驾驶，无需人类干预] ```

1.1.1 Key Definitions and Common Misconceptions

Each level of the SAE framework has specific, testable requirements that eliminate ambiguity for consumers and regulators:

• **Level 0 (No Automation):** The human driver performs all dynamic driving tasks. Examples include 20th-century vehicles with no driver assistance features, and even modern vehicles with basic backup cameras, which only provide information but do not control the vehicle.
• **Level 1 (Driver Assistance):** The system controls either steering or speed/longitudinal control, but not both simultaneously, and the human driver must monitor the driving environment at all times. The most common example is adaptive cruise control (ACC), which adjusts vehicle speed to maintain a safe following distance from the car ahead, but requires the human to steer.
• **Level 2 (Partial Driving Automation):** The system controls both steering and speed/longitudinal control simultaneously in limited scenarios, but the human driver is *legally and operationally responsible* for monitoring the driving environment and intervening if the system fails. The most widely deployed Level 2 systems include Tesla Autopilot, Ford BlueCruise, and GM Super Cruise, which can steer, accelerate, and brake on controlled-access highways, but require the driver to keep their hands on the wheel and eyes on the road at all times. A common misconception is that Level 2 systems are “autonomous”; under SAE definitions,