LiDAR was once regarded as an indispensable 'safety crutch' for autonomous driving, but solutions represented by Tesla have showcased the potential of pure vision. The vision-based approach not only offers cost advantages but also provides a solution closer to general artificial intelligence in simulating human driving behavior, handling complex semantic information, and ensuring consistency in system decision-making.

Is the logic more human-like?

Human drivers rely solely on optical information from their eyes, combined with logical reasoning in the brain, to navigate extremely complex traffic environments. Proponents of the pure vision approach argue that cameras, as the core carriers of perception, provide far richer information than LiDAR.

While LiDAR excels at precise ranging, it is essentially 'color-blind' and cannot recognize the color of traffic lights, text on road surfaces, or textures of object surfaces. In contrast, high-resolution images captured by cameras contain deep semantic dimensions. The system can not only detect obstacles ahead but also determine their identities through visual features, such as identifying whether a figure by the roadside is a child playing and likely to dart into the road or merely a plastic bag blowing in the wind.

This deep understanding of semantic information endows the vehicle with stronger predictive capabilities. Traditional LiDAR-based solutions may fail to detect or misjudge unusual objects or long-tail scenarios due to the lack of corresponding geometric models in their databases.

In contrast, vision systems, through deep learning, imitate human cognitive patterns and can learn the behavioral patterns of different objects. After identifying an object's category and state, the vision system can predict its next move based on experience. This 'decision-making intelligence' based on semantic understanding is difficult to achieve through point cloud ranging alone.

Meanwhile, with advancements in high-definition CMOS technology, vision-based solutions have demonstrated the potential to catch up with or even surpass LiDAR in long-range perception and spatial continuity.

Decision-making conflicts among sensors and the cost of redundancy

Many believe that more sensors equate to greater safety, but multi-sensor fusion can sometimes lead to fatal 'decision conflicts.' Due to the differing physical characteristics of cameras, LiDAR, and millimeter-wave radars, their perception results for the same environment often deviate.

For example, when millimeter-wave radar falsely detects an obstacle due to bridge reflections while the camera shows clear road conditions, the system faces a dilemma of whom to trust—a common cause of 'phantom braking.'

Rather than struggling with painful and potentially erroneous logical trade-offs among multiple imperfect sensors, the pure vision solution focuses all computational power and research efforts on the richest data source, eliminating uncertainty through algorithmic purity.

Beyond decision-making risks, LiDAR is not omnipotent in practical maintenance and environmental adaptability. While its active ranging capability is favored, laser beams scatter in heavy rain, dense fog, or blizzards, leading to sharply reduced detection accuracy or a flood of false signals.

LiDAR's hardware structure is extremely precise and fragile, making it vulnerable to damage when mounted on rooftops or similar locations. Post-installation calibration and cleaning maintenance costs are significantly higher than those for cameras. For automakers pursuing large-scale production, these high hidden costs and hardware complexity, to some extent, limit the Popularization speed (popularization speed) and lightweight evolution of autonomous driving system architectures.

Does the rise of occupancy networks make pure vision more promising?

Historically, the biggest weakness of pure vision solutions was depth perception—the inability to directly provide precise distances like LiDAR. However, with the introduction of occupancy networks and Bird's Eye View (BEV) technology, this challenge is being algorithmically reconstructed.

BEV technology unifies 2D images from multiple cameras into a 3D top-down view, eliminating judgment errors caused by perspective occlusion. Occupancy networks take this further by abandoning the need for object labeling. Instead, they divide surrounding space into tiny 'voxels,' directly predicting whether each spatial unit is occupied and its motion trend.

This shift from 'feature engineering' to 'spatial reconstruction' gives the vision system a 'spatial intuition' similar to LiDAR, enabling safe avoidance of unfamiliar obstacles by sensing physical occupation.

These algorithmic advancements blur the boundaries between perception and planning. By incorporating temporal information, the system can 'remember' previously seen scenes and predict future changes, greatly enhancing its handling capabilities at complex intersections and in human-vehicle mixed scenarios.

More importantly, the geometric representations output by occupancy networks can directly interface with planning models, generating smoother vehicle trajectories that align with physical laws and human driving habits. When computational power and model precision reach a critical point, the pure vision solution effectively replaces rigid hardware stacking with more flexible software complexity.

The upper limit of data-driven closed loops and end-to-end intelligence

The ultimate competition in autonomous driving lies in data. The low cost of pure vision hardware allows for effortless deployment across hundreds of thousands or even millions of mass-produced vehicles, creating an extremely vast data collection network. Whenever vehicles encounter edge cases on real roads, this data is transmitted to the cloud, continuously feeding algorithmic models through automated labeling techniques.

This scalability effect is difficult for LiDAR-based solutions to match. With exponential growth in training data, the generalization capabilities of pure vision systems become extremely powerful, enabling them to eliminate reliance on high-definition maps and achieve truly universal autonomous driving.

Currently, the autonomous driving industry is undergoing a comprehensive transition toward 'end-to-end' (End-to-End) large models. This architecture integrates perception, prediction, and planning into a unified neural network, where raw video streams are inputted to directly output vehicle control commands.

End-to-end models learn driving skills by imitating massive amounts of high-quality driving data, much like experienced human drivers, rather than memorizing rigid rule-based code. Since the core of such models is processing visual information streams, they naturally align with the pure vision approach.

When a sufficiently intelligent 'brain' can perfectly parse physical laws from video sequences, the ranging information provided by LiDAR becomes increasingly marginal. By pursuing the upper limits of artificial intelligence, the pure vision solution aims to achieve more scalable intelligent evolution on a simpler hardware foundation.

-- END --