Tesla AI director reveals the pure vision system: radar is just a crutch

Pay attention to the video number, find more exciting In the process of developing Tesla’s autopilot function, Elon Musk did not use on-board radar. A few days ago, the head of Tesla’s artificial intelligence explained the reason at a meeting. He said that the vision system developed by Tesla can completely replace other sensors. Electric car manufacturer Tesla has its own unique way of developing autonomous driving functions. In a speech at the Computer Vision and Pattern Recognition Conference, Andrej Karpathy, Tesla’s senior director of artificial intelligence, explained the specific reasons. The current autonomous driving function is actually the sum of all personal safety systems in today’s cars. Many new cars are equipped with functions such as automatic braking, adaptive cruise control, lane keeping, and prevention of accidental accelerator pedaling. The main challenge facing automakers is to integrate all these functions into a system that is more effective and reliable than humans. The first problem is perception. The vehicle must be able to understand the surrounding conditions in order for these subsystems to function at the right time. Vehicle perception is also the biggest difference between Tesla and other automakers. Musk recently chose not to install radar sensors in Tesla cars, but chose to install only optical cameras to sense changes in the surrounding environment. All other automakers use optical cameras, on-board radars, and lidars to collect environmental data. Because lidar is expensive, it is relatively rare; and car manufacturers have widely adopted low-cost vehicle-mounted radars. “The radar is just a crutch,” Andrej Karpathy said at the conference. Automakers usually use on-board radar to measure relative speed and distance, but Tesla believes that optical cameras can also do the same. The difference between the on-board radar and the optical camera actually affects the perfection of Tesla’s autopilot function, so Tesla decided to focus on improving the camera system. Capaci said: “The vision system we have built in the past few years is very good and can leave other sensors far behind.” After perceiving or collecting environmental data, it starts to calculate and make decisions. Most of the content of Capaci’s speech was about Tesla’s artificial intelligence and computers. These sound quite complicated. If you are not familiar with professional terms such as three-dimensional convolution, sensor fusion, neural network, etc., then you have no way to understand Tesla’s decision-making methods. Capaci did not explain what Tesla has achieved compared to other companies in the field of autonomous driving. Auto parts and technology provider Aptiv has already traveled billions of kilometers in actual and simulated driving through advanced safety systems. Aptiv also has thousands of safety engineers and can directly provide any car manufacturer with an assisted driving system similar to Tesla’s Autopilot. But the company rarely talks about details. However, Tesla may have the advantage of scale. The company already has millions of electric vehicles available for testing, and vehicles that have already been sold are constantly learning to test. Capaci said that Tesla’s neural network works in a “shadow mode” on Tesla vehicles, predicting what the driver will do in the background, and enhancing the robustness of the autonomous driving software. Tesla CEO Musk said in the third quarter of 2020 that fully autonomous driving technology will achieve “the largest asset value increase in history.” For example, he said, self-driving cars can drive other people around when the owner is working or sleeping, so as to make money for the owner. Recent analysis estimates that Tesla’s current stock price is US$650, of which US$100 to US$200 are based on Tesla’s expectations that it will provide drivers with a higher level of autonomous driving functions before the end of this year. If Tesla can make real progress in demonstrating the safety of the autonomous driving system, other auto companies may compete to imitate Tesla and promote their own autonomous driving features. A huge innovation field and potential source of profit will emerge from this. However, investors seem to be waiting for concrete progress at the moment. Since the beginning of this year, Tesla’s stock price has fallen by about 12%, lagging behind the increase in the S&P 500 and Dow Jones Industrial Average