Musk Fires Again, Did Laser Radar Get Added Wrong?

Musk has once again "fired" at lidar.

In an interview, Uber CEO Dara Khosrowshahi shared his views on the "autonomous driving race," stating that "the lidar solution is more feasible in the short term."

Musk refuted this on social media, saying that lidar makes autonomous driving more dangerous.

Elon Musk specifically gave the example that Waymo is unable to operate on highways due to the uncertainty in sensor fusion, and stated that Tesla's safety improved after removing radar, so a pure vision-based approach is the key.

This is not the first time Musk has openly rejected the multi-sensor fusion approach, and it certainly won't be the last.

Can Musk's choice represent the correct answer for the industry?

Route Dispute

As early as 2015, Musk made clear his pursuit of a pure vision-based technology path. He believed that "LiDAR is expensive and unnecessary," "LiDAR is a crutch that will drive the company into a dead end," and "humans don't rely on LiDAR to drive, so AI doesn't need it either."

Apart from cost, the core reason Elon Musk rejects LiDAR is that he believes:

When fusing information from different sensors, deviations may occur, leading to safety risks. Ultimately, the triumph of autonomous driving will belong to the principle of ultimate simplicity: bionics inspired by the human eye.

With this attitude, in May 2021, Tesla's first batch of mass-produced radar-free vehicles rolled off the assembly line. Starting in 2022, all Model 3 and Model Y vehicles delivered in North America are no longer equipped with millimeter-wave radar, and the relevant entries have been simultaneously removed from the official website’s specifications.

In the same year, Tesla temporarily "added back" one 4D millimeter-wave radar in China to comply with updated regulations, but it has now returned to a configuration of 8 cameras.

The dispute between the vision-based faction and the LiDAR-based faction has never ceased domestically.

For example, Li Xiang, the founder of Li Auto, once said that if Elon Musk had ever driven on a Chinese highway late at night, Tesla would have kept lidar.

Huawei's Yu Chengdong also stated that Huawei insists on using LiDAR because it can improve safety.

He Xiaopeng, who insists on the visual approach, stated that the visual upper limit far exceeds that of LiDAR, and with the maturity of technology and increased computing power, the potential of pure vision will be greatly unleashed.

He believes that in the future, visual systems will be able to recognize small objects on the road such as nails and displaced manhole covers, whereas LiDAR relying on point cloud imaging has inherent limitations in perceiving such two-dimensional surface details that contain almost no height information.

However, I personally have some doubts about this kind of statement.

The LiDAR solution is also equipped with cameras; it does not rely solely on LiDAR. If future technology enables the system to identify objects like nails using cameras, wouldn’t a LiDAR solution integrated with cameras be able to achieve the same capability?

In short, Xiaopeng Motors' entire lineup does not come equipped with LiDAR, and even at the L3 stage aimed at regulatory implementation, it will continue to use a "pure vision" solution without employing LiDAR.

Horizon Robotics' latest HSD also follows the visual technology route, but we still see LiDAR on vehicles equipped with Horizon's solution.

This is because this lidar is only used for detecting irregular obstacles and other non-standard objects, in order to enhance active safety capabilities.

When we cover the LiDAR on the car, the navigation assistance system can still be used.

Moreover, it is reported that Huawei also has a similar direction in technology preview.

This can be considered a "compromise solution" between the two current approaches.

The two paths reflect different companies' thinking on the implementation paths, cost control, and safety issues of autonomous driving.

Head-to-Head Confrontation

Currently, the price of domestic models equipped with LiDAR has dropped to the 120,000 RMB range, while models with vision-based solutions are now priced as low as 70,000 RMB.

This means that the once highly-regarded assisted driving feature is rapidly becoming the "standard" for a new generation of smart cars.

On high-speed NOA, whether equipped with LiDAR or not, vehicles can complete automatic ramp entry and exit, active lane changing and overtaking, and adjust speed according to speed limits. The experience of mainstream products has already shown a trend of "little difference."

The true watershed appears in extreme scenarios and the lower limits of user experience, which is precisely the core of the debate between the LiDAR advocates like Li Xiang and Yu Chengdong, and the vision proponents like Elon Musk and He Xiaopeng.

In order to improve the performance of vehicles in complex road conditions and dangerous scenarios, lidar not only follows up with large models but also repeatedly refreshes the perception capabilities of lidar.

Huawei's LiDAR can achieve an effective sensing range of 300 meters and can operate in harsh lighting conditions. They have also developed an in-cabin LiDAR solution that ensures all-weather anti-interference capabilities.

Musk and He Xiaopeng are rapidly increasing computing power to push the limits of vision through software development.

Currently, Tesla's total cloud computing power has reached 100 EFLOPS, the highest among global automakers.

Xpeng Motors is the domestic brand with the largest model computing power, with a total cloud computing power reaching 10 EFLOPS. On the vehicle side, by equipping with three Turing chips, the computing power reaches 2250 TOPS, making it the passenger car with the highest computing power globally.

Interestingly, two seemingly opposing paths may eventually converge.

With the decline in the price of LiDAR, cost is no longer the key point of contention between the two sides; the only remaining issue is whether more efficient and safer technology actually requires LiDAR.

Automakers that insist on multi-sensor fusion also evolve towards data-driven AI algorithms. The existence of LiDAR is to provide nourishment for the "super brain." If one day the perception capability of cameras satisfies these engineers, then perhaps the "super brain" will no longer need LiDAR.

Just like the debate between pure electric vehicles and extended-range technology, for us it has never been about "who is right or wrong," but rather "who's experience we are more willing to pay for."

In conclusion

Musk's "pure vision belief" is the confidence to use the lowest hardware costs to build the highest technological barriers through algorithms and data. Rather than calling it a technological ideal, it is more of an ultimate internet business model: driving marginal costs close to zero and achieving ultimate profitability through scale.

The so-called "LiDAR hazard" may simply be an attempt to convince the public and capital markets that the disruptive path chosen by Tesla—emphasizing software over hardware—is the only right path for the future.

Before the "standard answer" for autonomous driving is revealed, sales have already proven that lidar still earns the most honest vote from users.