The central dispute in autonomous driving perception: should systems use LiDAR or pure cameras to “sense” the road environment? The debate has simmered since 2021 when Elon Musk publicly called LiDAR “a fool’s errand,” and remains unresolved.
## LiDAR: Principles and Advantages
LiDAR (Light Detection And Ranging) emits laser pulses and measures reflection time to generate precise 3D point clouds. It can generate hundreds of thousands of 3D points per second, unaffected by lighting conditions, offering dramatically superior reliability in nighttime, tunnel, and low-light scenarios versus cameras.
Key suppliers: Hesai Technology (Nasdaq-listed, China’s market leader), RoboSense, Innovusion, and Velodyne (US, merged with Ouster). LiDAR costs have dropped from $100,000/unit in 2016 (Velodyne HDL-64E) to $200-500/unit in 2024 (Hesai AT128), enabling automotive-grade mass production. Xpeng X9 uses three LiDAR units; Leapmotor C10 uses one; NIO ET7 uses one 1550nm LiDAR (long range, eye-safe).
## The Pure Vision Logic
Tesla’s core argument: humans drive with only two eyes (cameras) — if AI is powerful enough, camera data theoretically contains all information needed for driving. Additional sensors cannot solve the fundamental problem, which is algorithms and data, not hardware.
Tesla’s weapon: real-world driving video data from 4 million global Teslas, continuously learning through shadow mode. FSD v12 adopted an end-to-end neural network that directly outputs steering angles and pedal values from raw camera inputs, bypassing intermediate rule sets.
## End-to-End Models: Blurring the Two Paths
Since 2024, “end-to-end autonomous driving” has become the field’s hottest topic. Tesla FSD v12, Xpeng XNGP 3.0, and Huawei ADS 3.0 all announced end-to-end architectures — merging perception, prediction, and planning modules into a single large model that maps directly from sensor inputs to control outputs.
Under the end-to-end framework, LiDAR data can equally serve as input, and the essential difference between the two paths is converging. The debate is shifting from “which sensors” to “how much data to train how large a model.”
See [Autonomous Driving Levels](https://sunqi.org/autonomous-driving-levels-en/) and [In-Vehicle LLMs](https://sunqi.org/car-large-language-model-en/).
