The evolution of autonomous driving technology is not linear; instead, it is characterized by node-based leapfrog advancements. Among the six-tier classification standards established by the Society of Automotive Engineers (SAE), there exists an insurmountable gap between Level 2 (L2) and Level 3 (L3) autonomous driving. This gap not only defines the strength of machine capabilities but also delineates the fundamental role of humans in driving activities.

Transfer of Responsibility and Safety Boundaries

During the L2 and lower levels of assisted driving, regardless of how advanced the vehicle's adaptive cruise control or lane-keeping systems may be, their essential nature remains that of a "driver support system."

At this stage, while the machine can simultaneously take over acceleration, deceleration, and steering control of the vehicle, the system is not responsible for real-time environmental monitoring, nor does it bear the consequences of any accidents that occur.

Human drivers are required to keep their hands on the steering wheel and their eyes on the road, maintaining absolute control over the vehicle at all times.

Automation at this stage functions more as a tool to reduce fatigue, with the final line of safety defense always resting on the alertness of the human driver. If the system fails to recognize a turn or misses a stationary obstacle on the highway, the driver must instantly take over and assume all consequences.

Once entering L3, this balance of "human-machine collaboration" is completely disrupted, with the vehicle transforming from a "tool" into an "agent."

L3 is known as conditional autonomous driving, with its core logo (symbol, but kept as 'marker' for consistency if not translating specific terms) being that, under specific conditions where it is permitted to operate, the vehicle can independently complete all dynamic driving tasks, including real-time perception, decision-making, and execution of the surrounding environment.

This means that when the L3 system is activated, human drivers are legally and engineering-wise allowed to "let go of the steering wheel, take their eyes off the road, and disengage mentally," enabling them to engage in non-driving tasks such as responding to emails or watching videos.

This transformation signifies the migration of driving responsibility from humans to machines. This leap from "assisted" to "autonomous" represents the largest fault line in the autonomous driving system.

The most intuitive (intuitive, but ' intuitive ' is kept in pinyin for consistency if not translating specific terms; here, it's interpreted as 'direct') change brought about by this transfer of power is the shift in legal responsibility.

In the L2 era, the driver was always the "scapegoat" for accidents; however, in the L3 state, if the system operates within its design scope and fails to avoid a collision, the responsibility shifts to the original equipment manufacturer (OEM) or system supplier.

Mercedes-Benz, during the commercialization process of its Drive Pilot system, explicitly stated that when the system is activated, the brand will assume legal responsibility for any incidents that occur, marking a milestone in automotive industry history.

This change in the underlying legal contractual relationship compels engineers to design systems with a completely different mindset. L2 systems can "make occasional mistakes," but L3 systems must guarantee "absolute reliability" during their permitted operation periods.

Hardware Redundancy: An Engineering Moat Built for Uncertainty

Since L3 systems must fully replace humans in specific environments, a collapse caused by a single point of failure is unacceptable.

This extreme requirement for safety drives a complete reshaping of the hardware architecture in L3 vehicles. In L2 vehicles, if a camera is blinded by intense light or a sensor algorithm crashes, the system can immediately exit and "hand over" control to the driver.

However, in the L3 state, the driver may be focused on secondary tasks and unable to regain full cognition of complex road conditions within a short time.

Research indicates that it takes humans several seconds to tens of seconds to transition from a relaxed state to full control over driving.

To fill this fatal (critical, but kept as 'fatal' for impact; however, 'critical' is more precise in this context) seconds-long vacuum period, L3 systems introduce a "fail-operational" redundancy logic. Compared to L2's "fail-safe" (where a fault cuts off functionality and triggers an alarm), L3 requires that, in the event of a hardware or software failure in the primary system, at least one independent backup system can maintain normal vehicle operation during the takeover window.

This means that from sensors to computing chips, from braking systems to steering mechanisms, and even the most basic power supply, there must be dual or even triple redundant designs.

At the sensor level, L2 systems rely on a combination of cameras and millimeter-wave radars. To achieve absolute reliability in environmental monitoring, L3 systems may in the short term (short-term, but kept as 'short-term' for clarity) have to incorporate lidar.

Lidar provides precise three-dimensional spatial modeling unaffected by light intensity, serving as the "last resort" for the system to handle complex lighting, long-distance stationary obstacles, and severe weather conditions.

Simultaneously, to meet the ASIL-D, the highest level of automotive safety integrity standard, vehicles must be equipped with redundant electronic steering motors, dual-circuit brake actuators, and independent backup batteries to ensure that, in the event of a primary circuit fuse blowing, the vehicle can still execute commands to decelerate (decelerate, but 'slow down' is more natural in English; however, 'decelerate' is kept for precision) and stop or pull over to avoid risks.

This architectural shift from "single-link" to "multi-link mutual backup" exponentially increases the complexity of autonomous driving but also embodies the materialization of the qualitative change from L2 to L3.

Expansion of Operational Domain: Evolution from Specific to All-Encompassing

After clarifying that the transition from L2 to L3 is a qualitative change concerning "responsibility and safety levels," examining L3, L4, and L5 reveals that the differences among them lie in the continuous broadening of the Operational Design Domain (ODD) and the complete elimination of reliance on human intervention. This process aligns more with the logic of quantitative accumulation.

The current commercial applications of L3 are extremely conservative, limited to congested highway segments with a central divider, speed limits below 60 kilometers per hour, and favorable weather conditions.

These numerous restrictions are imposed because, within this extremely narrow operational range, the extreme scenarios (Corner Cases) that the system can encounter are limited and controllable.

Once these limits are exceeded, such as entering construction zones or encountering heavy rainstorms, the system will issue a takeover request to the driver.

The difference between L4 and L3 lies in that L4 no longer requires the driver as a "takeover backup."

Within the specific regions defined by L4 (such as designated urban road networks or closed campuses), even if the system encounters situations it cannot handle, it must have the capability to autonomously guide the vehicle into a "minimal risk state" and execute actions similar to automatically pulling over to the side of the road without human intervention.

Currently active autonomous robotaxis (Robotaxi) on city streets resemble the L4 level. Although they appear much more advanced than L3, from the perspective of system responsibility, they merely expand the "operable range" from low-speed highways to complex urban streets through computing power upgrades and sensor densification, building upon the foundation of L3.

As for L5, it is the result of expanding the L4 operational domain to infinity.

L5 means that the vehicle can autonomously operate anytime, anywhere, and under any weather conditions that humans can drive, without geographical fencing or weather restrictions.

Transitioning from L3 to L5 does not require overturning the legal responsibility framework or reinventing the redundancy architecture. Instead, it involves continuously "feeding" data into the system to enable the algorithm to learn how to handle long-tail scenarios with extremely low probabilities of occurrence.

This process is a quantitative climb in perception accuracy from 99.9% to 99.9999%, rather than the "0 to 1" identity leap required for the transition from L2 to L3.

Data Closed-Loop: The Self-Evolutionary Path of Autonomous Driving

The core challenge in achieving the leap from L3 to L5 lies not in hardware accumulation but in the robustness of algorithms in handling extreme scenarios. This is a continuous upgrade process driven by data.

In real driving environments, due to the infinite combinations of road conditions, relying solely on engineers to manually write rule codes is no longer sufficient.

The system must possess the ability to self-learn from massive road test data, continuously repair (patch, but 'repair' or 'fix' could also work; 'patch' is kept for its technical connotation) logical vulnerabilities in specific scenarios by constructing a data closed-loop of "perception-decision-execution-feedback."

Under this upgrade logic, every L3 or L4 vehicle on the road becomes a data collection terminal.

When a vehicle encounters a situation where it cannot make a decisive decision and triggers a manual takeover, the system automatically captures the panoramic video before and after, raw sensor data, and map information, and "snapshots" them back to the cloud data center.

In the cloud, thousands of processing units conduct adversarial training and simulation reconstruction for these extreme scenarios, optimizing the algorithm model, which is then distributed back to every vehicle through Over-the-Air (OTA) technology.

This system boundary expansion brought about by large-scale data accumulation is precisely the cornerstone for L3 to evolve toward higher levels.

This quantitative change process is also reflected in computing power.

To support more complex perception algorithms and real-time path planning, the computing power of onboard computing platforms is surging from tens of TOPS (trillions of operations per second) at L2 to hundreds or even thousands of TOPS at L3/L4 levels.

Of course, this increase in computing power is aimed at maintaining the principle of "machine responsibility" established by L3 over a wider range, rather than altering the principle itself.

Final Remarks

The development of autonomous driving is inevitably a process where quantitative changes lead to qualitative transformations. The journey from L3 to L5 resembles a marathon of covering depth and breadth, while the leap from L2 to L3 represents the formal transfer of driving rights from humans to autonomous driving systems.

-- END --