特斯拉最新發布的自動駕駛系統引起了很多人的興趣，但人們更應該關注的，其實是該系統背后的若干新型技術。

特斯拉CEO埃隆?馬斯克在一個活動上介紹說，這套最新的自動駕駛系統，之所以具備了持續學習和改進的能力，是得于機器學習算法、汽車以及特斯拉車型的無線聯網功能，以及加上特斯拉所采集的詳細的測繪和傳感器數據。

機器學習算法是計算機科技領域的最新進展，計算機可以對海量數據進行分析，并利用它做出越來越精準的預測。簡言之，就是機器具有了學習功能。現在像谷歌、Facebook和特斯拉等企業現在都在利用機器學習技術“訓練”軟件，為客戶提供幫助或向他們銷售新的服務。

機器學習技術是計算機走向人工智能的方法，同時也是人工智能的一種形式。雖然馬斯克本人就人工智能的危險曾經發表過不少聳人聽聞的言論，不過他在上述活動上澄清說，他只是擔心那些被用作邪惡用途的人工智能。

在媒體提問環節，有位記者問道，特斯拉的自動駕駛服務與市面上其他基于計算機技術的駕駛輔助功能有什么區別時，馬斯克強調了特斯拉自動駕駛系統的學習功能。

他表示：“整個特斯拉車隊是在同一個網絡上操作的。當一輛車子學到了什么東西，所有的車子也就都學到了。這是就超越了其他汽車公司做不到所做的。”在講解自動駕駛軟件時，馬斯克解釋道，每名利用自動駕駛系統進行駕駛的駕駛員，實質上都變成了一名“教自動駕駛系統應該如何工作”的專家培訓師。

Tesla’s new autopilot system is relying on the cutting edge of machine learning, connectivity and mapping data.

While Tesla’s new hands-free driving is drawing a lot of interest this week, it’s the technology behind-the-scenes of the company’s newly-enabled autopilot service that should be getting more attention.

At an event on Wednesday Tesla’s CEO Elon Musk explained that the company’s new autopilot service is constantly learning and improving thanks to machine learning algorithms, the car’s wireless connection, and detailed mapping and sensor data that Tesla collects.

Tesla’s cars in general have long been using data, and over-the-air software updates, to improve the way they operate.

Machine learning algorithms are the latest in computer science where computers can take a large data set, analyze it and use it to make increasingly accurate predictions. In short, they are learning. Companies like Google GOOG 0.07% , Facebook FB 1.63% and now Tesla TSLA 2.58% are using machine learning as a way to train software to help customers or sell them new services.

Machine learning is the way that computers can become artificially intelligent, and the technology is a form of AI. While Musk has taken a sort of alarmist stance against the dangers of AI, he clarified during the event on Wednesday that he’s only concerned with artificial intelligence that is meant for nefarious purposes.

When a reporter asked Musk during the media Q&A what made his company’s autopilot service different than other computer-based driving assistance features that competing big auto makers are working on, Musk emphasized learning.

“The whole Tesla fleet operates as a network. When one car learns something, they all learn it. That is beyond what other car companies are doing,” said Musk. When it comes to the autopilot software, Musk explained that each driver using the autopilot system essentially becomes an “expert trainer for how the autopilot should work.”

也許大多數汽車公司都沒有打造自己的機器學習系統，但谷歌的自動駕駛汽車技術卻與特斯拉有異曲同工之妙。

從某種角度上看，特斯拉出品的汽車相比于傳統汽車，特斯拉汽車倒是更像是一種更類似于智能聯網產品，比如說智能家居產品公司Nest的學習型恒溫器（現在由谷歌母公司旗下的Alphabet公司所有）。Nest公司的恒溫器采用了傳感器和計算機算法，能夠逐漸學習使用者的行為模式，并能通過軟件升級提供更加有用的服務，甚至能影響Nest 生產什么樣的下一代硬件的生產決定。

那么，特斯拉的自動駕駛系統乃至車子本身是如何學習的？一切都從數據開始。

構建這種類型的駕駛員輔助系統的企業，包括像谷歌這樣正在設計完全自動駕駛技術的公司，都需要“教”會計算機如何利用數據傳感器系統，而不是人類的感觀，來接管汽車駕駛的核心部分。要達到這個目標，這些企業首先要利用大量的數據來對計算機算法進行“訓練”。

英偉達公司（Nvidia）的汽車部門高級總監丹尼?夏皮羅在接受《財富》（Fortune）采訪時表示，你可以把它想象成一個孩子通過經驗和重復來不斷學習的過程。針對自動駕駛汽車和駕駛員輔助應用，英偉達推出了一系列能讓電腦處理海量數據的高性能芯片，最近還推出了一個名叫Drive PX的計算機系統。

為了構建一輛自動駕駛汽車，汽車公司需要把長達幾十萬甚至幾百萬英里的駕駛視頻和數據輸入計算機的數據模型，從基礎上搭建一個龐大的駕駛“詞庫語境”。計算機算法會利用虛擬技術分析和理解這些視頻。其這樣做的目的是，當意外事件發生時，——比如一個球滾到了路上，——汽車可以自動識別當前模式，并采取相應規避措施（減速行駛，因為可能會有小孩跑到馬路上撿球）。

英偉達所做的，則是把這個龐大的“駕駛辭典”載入一個強大而小巧的計算機硬件里，使它能夠被用到汽車上。此后，包括谷歌和特斯拉在內的一些公司則會通過多種來源繼續添加各種類型的數據，讓模型繼續進行學習。

要使汽車計算機能夠在馬路上做出更智能的決策，汽車公司會盡可能收集更多的數據，其中既包括顧客的駕駛數據，也包括GPS和地圖數據，以及企業員工駕駛測試車時產生的數據。

While most car companies might not be building learning systems, Google’s self-driving cars operate in a similar manner.

In that way, Tesla’s cars are more similar to smart connected gadgets like Nest’s learning thermostat (now owned by Google’s Alphabet), than they are to traditional cars. Nest’s thermostat, using sensors and algorithms, learns its owner’s behavior over time, and through software updates offers increasingly useful services, or even informs Nest’s decisions about its next-generation of hardware.

So, how does Tesla’s autopilot system, and its cars in general, learn? It all starts with data.

Companies building these types of driver-assistance services, as well as full-blown self-driving cars like Google’s, need to teach a computer how to take over key parts (or all) of driving using digital sensor systems instead of a human’s senses. To do that companies generally start out by training algorithms using a large amount of data.

You can think of it how a child learns through constant experiences and replication, explained Nvidia’s Senior Director of Automotive, Danny Shapiro in an interview with Fortune. Nvidia NVDA 1.57% sells high performance chips that enable computers to process large amounts of data, and more recently started selling a computing system, called Drive PX, for self-driving cars and driver-assist applications.

To create a self-driving car, companies feed hundreds of thousands, or even millions, of miles of driving videos and data into a computer’s data model to basically create a massive vocabulary around driving. The algorithms use visual techniques to break down the videos and to understand them. The goal is that when something unexpected happens — a ball rolls into the street — the car can recognize the pattern and react accordingly (slow down because a child could be running into the street after it).

For Nvidia, the company loads this “driving dictionary,” as Shapiro calls it, onto powerful but compact computing hardware that can be used on the car. After that, companies like Google and Tesla add various types of data from different sources to continue to inform the model over time.

Companies try to gather as much data as possible to help a car’s computer make smarter and better decisions on the roads. This includes data from customers driving, data from GPS and maps, and data from company employees driving research cars.

為了更好地讓自動駕駛系統進行學習，特斯拉正在制作詳細的高清地圖。另外，借助汽車歸功于其車型的硬件配置，特斯拉還能采集駕駛員的數據。去年生產的所有特斯拉汽車均生產的所有轎車在底部都安裝了12個顆傳感器，后視鏡附近裝有一個顆前視攝像頭，車底前部還安裝了一套雷達系統。這些傳感系統源源不斷地收集著數據，既幫助了特斯拉的自動駕駛技術走向成熟，也使特斯拉未來的駕駛體驗能變得更好。

另外，所有特斯拉汽車都配備的車子都安裝了常年開啟的無線連接裝置。在這些裝置的幫助下，，因而自動駕駛系統的駕駛信息和使用信息經收集后，會自動發送到云端，然后被軟件分析。此外特斯拉還收集了那些搭載新型自動駕駛技術和車道轉換系統的汽車的信息，并用這些信息對算法進行訓練。然后，特斯拉會對這些算法進行路試，并將它們整合到即將推出的軟件里。

由于車企的具體目標不同，他們往往會依賴不同類型的數據。比如谷歌在自家的自動駕駛汽車上采用了又大又昂貴的LIDAR（激光雷達）系統。馬斯克認為，就對于特斯拉自動駕駛汽車的需求而言來說，LIDAR系統有點殺雞用牛刀的感覺。

不過馬斯克也表示，特斯拉的自動駕駛和車道轉換技術需要比標準導航技術更詳細的高清地圖數據。為了滿足這一需求，特斯拉已經開始打造自己的高清地圖——其精細度要高于標準導航系統100倍。它的數據大多數來自路面上行駛的特斯拉轎車，但也有一些是特斯拉員工駕駛測試車所獲得的數據。

這些新的服務也為企業帶來了意想不到的業務模式。馬斯克表示，特斯拉將來可能會有興趣把這些地圖數據賣給其他汽車公司。

特斯拉并不是唯一一家研究汽車輔助和自動駕駛技術的公司。谷歌在未來科技方面一直走在前面，奧迪（Audi）也有自己的交通擁堵輔助軟件。英偉達公司的夏皮羅表示，目前大多數汽車制造商都在研究這些技術。

英偉達的Drive PX系統已于今年夏天開始出貨。夏皮羅表示，英偉達已經與50多家公司和研究機構進行了接洽。特斯拉Model S的17寸大屏和儀表盤就使用了的就是英偉達的芯片。，目前也有人推測稱，特斯拉在Model X SUV的未來版本中可能將使用Drive PX系統。不過夏皮羅并沒有詳細討論英偉達與特斯拉或奧迪等公司的關系的細節（奧迪的在其交通擁堵輔助系統里也使用了英偉達的技術）。

夏皮羅最后指出，雖然有些公司已經開始采用這些技術，但目前仍是汽車自動駕駛技術的早期階段。“未來10年還需要完成，這方面將有大量的工作會完成。”（財富中文網）

譯者：樸成奎

審校：任文科

The data from Tesla drivers was enabled by the hardware choices that Tesla has made. All Tesla cars built in the past year have 12 sensors on the bottom of the vehicle, a front-facing camera next to the rear-view mirror, and a radar system under the nose. These sensing systems are constantly collecting data to help the autopilot work on the road today, but also to amass data that can make Tesla’s operate better in the future.

Because all of Tesla’s cars have an always-on wireless connection, data from driving and using autopilot is collected, sent to the cloud, and analyzed with software. For autopilot, Tesla takes the data from cars using the new automated steering or lane change system, and uses it to train its algorithms. Tesla then takes these algorithms, tests them out and incorporates them into upcoming software.

Companies will rely on different types of data depending on what they’re trying to do with the cars. For example, Google has used large and expensive LIDAR (light-based radar) sensors on its self-driving cars. But Tesla’s Musk said that LIDAR was basically overkill for what Tesla’s autopilot cars need.

But Musk said that Tesla wanted much more detailed high-precision mapping data for its automated steering and lane change applications than was available through the standard navigation tech. To meet its needs, Tesla has started to build high-precision maps —that have 100 times the level of granularity compared to standard navigation systems — using mostly data from Tesla cars driving on roads, but also some data from Tesla employees driving research cars.

These new services could provide unexpected business models for companies. Musk said that Tesla might be interested in selling the mapping data to other car companies down the road.

Tesla isn’t the only car maker working on driver-assist and self-driving car tech. Google is blazing ahead on its futuristic tech, while Audi has traffic jam assist software. Nvidia’s Shapiro says that most automakers are investigating these technologies.

Nvidia started shipping Drive PX this summer, and Shapiro says that it’s engaged with over 50 companies and researchers. Tesla uses Nvidia chips in the 17-inch screen and the instrument cluster for its Model S and there has been speculation around whether Tesla might use the Drive PX system in future versions of the Model X SUV. Shapiro wouldn’t discuss the specifics of its relationships with Tesla or Audi, which uses Nvidia’s tech in its traffic jam system.

Shapiro cautioned that despite some companies already deploying these technologies, it’s still early days for self-driving car tech. “A huge amount of work will be done on this over the next decade,” he said.