[Introduction] As of late April 2026, the autonomous driving industry is undergoing a dual inflection point: On one side, the Beijing Auto Show has sparked ongoing debates between the L3 and L4 routes. Huawei and Geely insist that L3 is an essential phase for the industry, while companies like XPeng and Pony.ai firmly advocate skipping L3 and moving directly to L4. On the other side, the Ministry of Public Security has officially released the industry standard Safety Traffic Norms for Road Testing and Demonstration Applications of Intelligent Connected Vehicles (GA/T 2388-2026), set to take effect on July 1. This standard establishes unified national traffic rules, compliance criteria, and liability boundaries for high-level autonomous driving at L3 and above.

This debate transcends mere technological route selection; at its core, it represents the industry's ultimate strategic game theory (can be translated as "game" or "competition") over commercial feasibility, safety logic, and regulatory adaptability. L3 is inherently paradoxical and serves only as a transitional product for the industry, while L4 autonomous driving, having undergone triple iterations in technology, data, and models, has moved beyond the pilot phase to become the sole viable, scalable, and sustainable commercial endpoint for autonomous driving.

L3's Theoretical Deadlock and Compliance Challenges Make It an Industry 'Transition Maze'

The core definition of Level 3 autonomous driving is that the system can independently complete all dynamic driving tasks, but the driver must remain on standby to respond to system handover requests and quickly intervene in vehicle control. While it appears to be a perfect transitional solution between L2 assisted driving and L4 fully autonomous driving, L3 has revealed fatal flaws in theoretical research, legal definitions, and practical implementation.

Recently, arXiv published the paper The L3 Impossibility Theorem: A Formal Analysis of Human-Automation Handover, which, from the perspective of formal human-machine interaction modeling, proves the theoretical limitations of L3: In complex dynamic road scenarios, humans cannot complete stable and reliable takeover operations within an extremely short timeframe. The 'human-machine co-driving, always ready to take over' model contains unavoidable safety vulnerabilities, which are the core reason for L3's consistently high accident rates.

At the legal and compliance levels, L3 is trapped in a dilemma of ambiguous rights and responsibilities. The Equipment Industry Development Center of the Ministry of Industry and Information Technology clarifies that the core difference between L3 and L2 is the shift in liability from humans to the system. However, the division of rights and responsibilities during handover moments remains in a gray area. In domestic L3 pilot scenarios, automakers face excessive liability risks, and the vehicle insurance pricing system is inadequate, resulting in low corporate willingness to implement L3 and slow commercialization progress.

The new regulations to be implemented by the Ministry of Public Security in July further amplify L3's awkward position. For the first time, these regulations quantify comprehensive driving standards for intelligent vehicles, including acceleration/deceleration, following distances, lighting usage, and emergency responses. They also clarify rules for determining violations and faults in intelligent vehicles, as well as accountability mechanisms. For L3 models requiring human-machine collaboration, the need to comply with stringent machine compliance standards while relying on unstable human takeover operations results in extremely low dual fault tolerance. Research and development costs, as well as compliance costs, double, yet user experience and commercial value see minimal improvement.

In short, L3 is not a stepping stone to advanced autonomous driving but rather a transitional solution characterized by technological redundancy, ambiguous rights and responsibilities, compliance difficulties, and weak profitability. It is trapped in an industry maze with no clear path forward.

L4 is Not a Leap of Faith but an Industry Return Driven by Technological Maturity

Compared to L3's inherent flaws, the core advantages of L4 autonomous driving are straightforward: Within its designed operational domain (ODD), the system independently completes all driving tasks without human driver intervention or handovers. It completely sheds the technological burdens, legal disputes, and safety hazards associated with human-machine transitions.

Since 2026, advancements in physical AI, closed-loop data from massive scenarios, the deployment of end-to-end large models, and the support of unified national regulations have enabled L4 to move beyond 'conceptual pilots' and reach a tipping point for large-scale deployment. A strategic consensus among industry leaders has become increasingly clear: Skipping L3 and moving directly to L4 is not a radical gamble but the optimal solution that follows industry regularity .

1. Technological Dimension: Physical AI Reconstructs Perception Logic, Achieving Full Compliance

Traditional autonomous driving relies on modular rule-based algorithms, which can only recognize known scenarios and fail to understand complex physical world logic. This is the core reason for previous incidents involving unmanned vehicles obstructing roads or mishandling emergencies. The emergence of physical world large models in 2026 has completely reconstructed the technological foundation of L4.

Companies like Momenta and AutoX unveiled their latest technological solutions at the 2026 Beijing Auto Show, achieving an upgrade from 'image recognition' to 'physical cognition.' These systems can independently judge object inertia, spatial relationships, and scene causality, predicting long-tail unexpected scenarios such as a football rolling onto the road or crowds crossing.

More critically, after the implementation of the Ministry of Public Security's new traffic norms, leading L4 solutions have completed underlying adaptations. They have incorporated six major compliance red lines—including turn signal delays, acceleration/deceleration thresholds, yielding priorities, and fault-induced roadside parking—into the fundamental logic of their algorithms. This completely resolve (can be translated as "completely resolves") the industry chaos of intelligent vehicles driving recklessly or failing during emergencies, achieving bidirectional alignment between technological capabilities and official compliance standards.

2. Data Dimension: Pure Machine Decision-Making Closes the Loop, Creating a High-Quality Data Moat

The upper limit of autonomous driving capabilities is fundamentally determined by high-quality scenario data. Under L3, vehicles are primarily driven by humans with system-assisted monitoring, collecting only human-machine hybrid driving data. This data fails to truly reflect the system's independent decision-making capabilities and holds extremely low value, making it difficult to support the iteration of high-level intelligent driving.

In contrast, L4 involves full machine autonomous decision-making. Every road condition response, braking maneuver, and route planning serves as a genuine and effective training sample, forming a complete data flywheel of 'road testing - iteration - deployment.' Leading players continuously accumulate massive amounts of data on urban commuting, highway travel, and mixed traffic in complex urban areas while leveraging simulation platforms to complete daily validation of millions of scenarios, addressing gaps in long-tail data.

As the industry consensus goes: L3 accumulates ineffective human-machine handover data. Only L4's pure machine decision-making data can support the continuous evolution of autonomous driving.

3. Model Dimension: End-to-End Architecture Lands, Eliminating Modular Technological Redundancy

Traditional autonomous driving relies on a layered modular architecture for perception, prediction, planning, and control, which suffers from severe information loss. These modules have poor adaptability and high algorithmic redundancy. The CVPR 2026 paper UniL4, published on April 25, 2026, introduced the first unified end-to-end framework for urban L4 driving, integrating full-chain functionality and achieving 92% autonomous mileage without human intervention in multi-city road tests.

The core value of end-to-end models lies in replacing tens of thousands of lines of manual rule-based code with a single AI model, simplifying the technological architecture, reducing failure probabilities, and accelerating decision-making. Tesla's FSD V12 and XPeng's latest intelligent driving systems adopt similar approaches, which is also the core technological confidence driving automakers to abandon L3 and focus on L4.

Full-Scale Commercialization of L4: Multi-Scenario Blossoming

The biggest market question (can be translated as "doubt") about L4 has been that 'the technology is mature, but commercialization remains distant.' However, in 2026, driven by cost reductions, regulatory compliance, and the land (can be translated as "implementation") of benchmark cases, L4 has formed four mature deployment scenarios: low-speed buses, robotaxis, urban logistics, and commercial vehicles in closed areas. It has moved beyond pilot hype to achieve sustainable commercial profitability.

1. Autonomous Driving Buses Go Global, Setting Public Transportation Benchmarks

Autonomous public buses represent the earliest scenario where L4 achieved large-scale and regularized operation. This field has moderate entry barriers, high social value, and strong replicability. Mushroom Autonomous Driving (Mushroom) stands out as the industry leader, creating domestic and overseas benchmark cases.

In domestic scenarios, Mushroom's front-mounted mass-produced L4 autonomous buses, MOGOBUS, have been deployed in over 20 cities across China, accumulating more than 5 million kilometers of safe driving. In March 2026, MOGOBUS launched the 'Qin-Ao Medical Line' in Hengqin, creating China's first cross-border autonomous microcirculation bus route for medical appointments, precisely addressing short-distance commuting and medical transportation needs in the area.

Overseas, Mushroom achieved a breakthrough for Chinese intelligent driving. In October 2025, Mushroom, in partnership with BYD, won Singapore's first official L4 autonomous bus project, marking the first entry of Chinese autonomous buses into a developed country's backbone (can be translated as "backbone") public transportation network. In April 2026, the vehicles arrived in Singapore, adapting to local left-hand driving and rainy, high-density road conditions. They are set to begin regular operations on Marina Bay (Binhaiwan) Route 400 and Weiyi Technology City (Wei Yi Tech City) Route 191 in the second half of the year, alleviating local bus driver shortages and validating L4 buses' global commercialization capabilities.

Compared to other scenarios, L4 autonomous buses combine public welfare attributes with commercial value, adapting to diverse scenarios such as urban microcirculation, tourist transportation, and cross-border travel. They are also the first L4 scenarios to pass official compliance and stability verification following the implementation of new regulations.

2. Robotaxi Costs Break Through Bottom Lines, Entering Profitability Cycle

As a core urban mobility scenario, the commercialization turning point for Robotaxi has arrived. Pony.ai publicly predicts that by 2027, the cost of fully autonomous Robotaxi vehicles will drop below 230,000 RMB, matching the price of ordinary passenger vehicles.

From a business model perspective, at this cost level, Robotaxis can achieve payback within 3-4 years and maintain stable profitability over an 8-10-year vehicle lifespan. Meanwhile, NVIDIA announced plans to launch large-scale unmanned mobility services at the 2028 Los Angeles Olympics, covering dozens of cities, signaling the formal large-scale commercialization of L4 passenger mobility.

3. Urban Last-Mile Logistics: A Low-Risk, Fast-Deployment Cash Cow Scenario

Unmanned delivery represents a pioneering track for L4 commercialization. Compared to passenger scenarios, freight logistics involves lower regulatory risks and simpler scenarios. Companies like JD.com and Meituan have deployed L4 unmanned delivery vehicles, achieving full-link delivery from warehouses to stores and stores to homes on urban roads. These vehicles adapt to diverse scenarios such as communities, industrial parks, and business districts, continuously reducing labor-based delivery costs and already achieving regional scalability and profitability. Neolix, a domestic leader in L4 urban last-mile logistics, completed multi-city large-scale deployment in 2026, establishing a standardized and replicable business model for unmanned logistics and becoming a benchmark for L4 last-mile logistics commercialization.

4. Closed-Scene Commercial Vehicles: Stable Deployment in Mines, Ports, and Industrial Parks

Closed and semi-closed scenarios such as airports, ports, mines, and industrial parks, with simple vehicle and pedestrian flows, serve as natural testing grounds and profitable scenarios for L4 autonomous driving.

The autonomous driving-specific traffic regulations set to take effect in July 2026 have completely end (can be translated as "completely put an end to") the industry's wild growth and clearly separated the winners from the losers: The inherent paradox of L3's human-machine co-driving cannot be resolved. Its high compliance costs and weak commercial value relegate it to a brief transitional phase.

In contrast, L4 autonomous driving, leveraging triple technological advantages—physical AI iteration, closed-loop high-quality data, and end-to-end model reconstruction—along with a unified national regulatory compliance system, covers all scenarios, including buses, taxis, logistics, and commercial operations. Benchmark cases such as Mushroom's overseas bus deployment and Pony.ai's cost reductions prove that L4 is no longer a distant future concept but a present-day, deployable, profitable, and replicable industry endpoint.

The competition in the autonomous driving industry has never been about 'gradual progress' through involution but about 'correct direction' for breakthroughs. Abandoning L3's ineffective maze and pursuing L4's endpoint track represents the sole optimal solution for the autonomous driving industry.