The issue of Tesla’s Autopilot abandoning LIDAR and using a pure vision solution has been hotly debated for more than three years, with Musk first suggesting that it could be done without LIDAR in 2018, saying “anyone who uses LIDAR is a fool” in 2019, and “pure vision FSD philosophy” on Tesla AI Day in 2021. By December 5, 2021, Tesla’s Fully Automated Driving FSD had been upgraded to version 10.6 Beta. Although several of these accidents have been questioned by the community and officials, and the technology of LiDAR is maturing and commonly used in almost all non-Tesla self-driving cars, and its price has dropped to hundreds of dollars, Musk and Tesla still insist on using the camera pure vision FSD, why?
This article on this from Tesla’s self-driving AI road test history, cost, AI core technology and marketing, and other aspects of a comprehensive comb.
1 Road test data: cumulative secret
Self-driving is actually AI driving, without a large amount of data for the algorithm to learn, AI can not evolve. Therefore, for autonomous driving manufacturers, road test data is a valuable and not easy to obtain.
In 2016, autonomous driving was still in a starburst, unlike the current blaze of fire.
For the full year 2016, Tesla sold 76,230 new vehicles worldwide, a 50.7 percent increase in sales compared to 2015.
Thanks to its self-selling vehicles and “crowdsourcing” collection strategy, Tesla is already way ahead of other automakers in terms of data accumulation. By the end of 2016, Tesla had acquired 3.5 billion miles of road test data.
Tesla has been able to obtain so much road test data, in addition to the data collection function of its own vehicles, its data-sharing strategy has also been very important in the privacy-conscious European and American markets.
Tesla’s data sharing strategy
Tesla updated its data sharing policy in early 2017, the main idea of which is to reach an agreement with vehicle owners to increase the amount of road test data of Tesla cars by collecting video from on-board cameras to pave the way for the technological development of autonomous driving.
Tesla officials said: “We’ve been working around the clock to improve the safety index of autonomous driving and bring it into everyone’s lives as quickly as possible. To make this happen, however, we need to collect small videos taken by cameras on the outside of the car to give us a clearer picture of road conditions such as lanes, street signs, and traffic signals.”
Tesla has written in an article, “If the faster we can understand these road conditions, there’s no doubt that your Tesla Autopilot will be more capable. In addition, we want you to be clear that to protect everyone’s privacy, your license plate number will never appear in a small video. It’s even more unlikely that we’ll let someone go through the Tesla system to search for video information about a specific vehicle.”
At the same time, Tesla also promises that even if the library does get hacked, it will be difficult for someone with ulterior motives to get too much relevant information from it. This data sharing policy will mainly apply to the Autopilot 2.0 system. The reason for this is that the detection suite installed on the new Autopilot system includes eight cameras, a full range of ultrasonic sensors, and a forward-looking radar. This means that the cameras will be able to capture more information and the amount of data will increase exponentially.
In addition to adding sensors and cameras to the car to collect more data in that data sharing strategy update in early 2017, Tesla Autopilot 2.0 was updated to include optimized features for HW2 hardware vehicles.
Having bridged the gap between Autopilot 2.0 and the previous generation of Autopilot, Tesla is on the fast track for road test data to be collected and used in the evolution of Autopilot.
In 2020, Tesla’s global ownership exceeded 1 million units, with 627,550 units delivered in the first three quarters of 2021 and total ownership expected to approach 2 million units in 2021.
It is estimated that a LIDAR-enabled autonomous driving test vehicle can generate up to 10 TB of data per day, and for pure vision FSDs, the raw data is much larger and the useful road test data obtained after cleaning can reach hundreds of megabytes to several gigabytes.
Such a huge amount of road test data is the “nutrient” for Tesla’s self-driving AI brain to evolve continuously.
2 The cost: is it really the lowest?
There has been a saying that one of the reasons Tesla Autopilot uses cameras is to reduce costs. From Autopilot 2.0, to the current FSD 10.6 Beta, Tesla models have been insisting on using eight cameras. Although there is no information available on the camera models in Tesla vehicles, the average price is expected to be around $10.
Now the cost of LIDAR has been reduced a lot to hundreds of dollars level, the current LIDAR route self-driving vehicles are generally equipped with a long-range LIDAR on the roof and a short-range LIDAR on each side. Its total price is about $600-800.
Therefore, Tesla pure vision FSD camera and lidar self-driving vehicle lidar cost difference is about: 700-10 * 8 ≈ 600 (U.S. dollars), which translates into about 4,000 yuan.
For car manufacturers, this cost difference should not be too big. However, autonomous driving requires a powerful AI brain in addition to a camera or LIDAR sensor. And the arithmetic power needed to process images and process LIDAR point cloud maps is very different, with the former being several orders of magnitude more than the latter.
This is one of the reasons why Tesla’s autopilot chip has a high arithmetic power.
Therefore, on balance, the claim that the cost of pure vision FSD solutions is lower is not valid.
3 Pure vision: FSD core technology
Tesla uses pure vision FSD solution, which is destined to require greater investment in AI processing of vision. The core technology of Tesla Autopilot mainly includes two parts, the HW series at the vehicle side and the AI supercomputing Dojo at the remote cloud side. Tesla’s HW series has undergone four generations of research and development.
One of the autonomous FSD autopilot chip launched in 2019, the arithmetic power can reach 144 TOPS, with this chip, Tesla cars can achieve better and reliable autopilot capability.
In 2020 Tesla released HW 3.0 based on the 14nm process and said that compared to the previous generation of Autopilot powered by Nvidia hardware, HW 3.0 is 21 times better in terms of frame rate while power consumption remains almost the same.
Tesla is already working on the next-generation chip HW4.0, which will be three times more powerful than HW 3.0 and is expected to start mass production in the fourth quarter of 2021.
AI Supercomputing Dojo
At the 2021 AI Day conference, Tesla presented its AI supercomputing Dojo. According to Ganesh Venkataramanan, head of the Dojo project, Musk asked Tesla engineers a few years ago to design an ultra-high-speed training computer, and that’s how Tesla started the Dojo project. the Dojo supercomputer will be commissioned next year to train AI algorithms based on a large number of videos.
Dojo is a distributed computing architecture connected by a network structure with a large computational plane, ultra-high bandwidth and low latency, large network partitioning and mapping, and more, with a new compiler to reduce local and global communication and scalability.
The supercomputing has Tesla’s own AI training chip, D1, built in. The D1 chip is manufactured using a 7nm process with a single chip area of 645mm² and contains 50 billion transistors, with peak arithmetic power of 362 TFLOPS for BF16/CFP8 and 22.6 TFLOPS for FP32, and thermal design power consumption (TDP) of no more than 400W.
Ganesh said that Tesla Dojo is the fastest AI training computer ever built. The Dojo supercomputers are 4 times more powerful than existing computers, 1.3 times more energy-efficient, and have a 1/5th the carbon footprint of the original.
Some say that Tesla CEO Musk’s adoption of purely visual technology is partly motivated by the marketing component.
In fact, rather than Tesla’s emphasis on promoting purely visual technology as a form of marketing, Tesla’s marketing is attributed to its CEO Musk’s personal IP marketing when it comes to putting life safety at stake in terms of autonomous driving.
According to statistics, in 2021 in China, Musk was on Weibo hot search 74 times a year, on average less than once a week (5 working days), and Musk’s image seems to be unaffected even in the face of a series of storms over rights protection. In fact, Musk is synonymous with Tesla’s enduring popularity.
Tesla has embarked on the road of pure visual FSD, and despite some accidents, its autopilot experience is currently more advanced overall. There is no doubt that Tesla will keep going on this road of pure visual FSD.
Of course, the autonomous driving route with LiDAR sensors has more manufacturers and more customers and consumer groups besides Tesla, and it is impossible to prove who can win more by using pure vision FSD or autonomous driving with multiple sensor fusion.
Perhaps, it is the same as the Apple iOS and Android camps in the smartphone market, and eventually, both have their own markets. And in the market, there will never be a second Apple, but there are countless Android manufacturers.