Tesla FSD's secret weapon: Dojo neural network

Publisher:trendsetter9Latest update time:2020-10-30 Source: EEWORLDKeywords:Tesla Reading articles on mobile phones Scan QR code
Read articles on your mobile phone anytime, anywhere

 (“Dojo” Nervous System) (“Dojo” Nervous System)

  

北京时间10月30日消息,据国外媒体报道,自特斯拉上周推出完全自动驾驶(FSD)测试版以来,获得测试资格的幸运儿发布的视频证实Autopilot有了显著的改善,曾被Autopilot 2.5视为挑战的任务如今已变得不再困难,而这一切背后的秘密武器则在于特斯拉即将推出的“Dojo”超级计算机。

  

Tesla owners who have used FSD for a while know that the Autopilot suite has some limitations, and everyone who has ordered FSD knows that it is a work in progress and that the feature will become more mature with the help of Tesla's neural networks. In the meantime, in order to process a larger amount of information and improve the vehicle's understanding of the surrounding environment, the Tesla AI team is developing Dojo. This is Tesla's neural network training program, which is designed to start breaking down 4D data instead of the "~2.5D" data that Autopilot used previously.

  

Musk gave a detailed introduction to the need to build a more complex autonomous driving system in the second quarter of 2020 earnings call:

  

"The big milestone that's really happening now is the shift in autonomous driving systems or automotive AI, from 2.5D to 4D. 2D is more like recognizing pictures that are strictly related in time and relatively isolated, while 4D is like a video that can be played continuously." The biggest problem with previous FSD and Autopilot versions was that the information transmitted through pictures was not enough. The understanding of the constantly changing surrounding environment requires time stamps and higher accuracy. The key is to transition from images or 2D as Musk said to video, which is 4D.

  

Musk pointed out that "the research we have been doing so far is really still focused on the 2D field, and although 2D has good temporal correlation, it is far inferior to a full 4D system. It is actually really hard to express in words how useful a full 4D system is. 4D systems have the ability to input isolated pictures instead of video in time, and the system can also turn pictures into surround video. Therefore, the car with 4D system will look much improved."


Roundabout navigation reflects the efficiency of the new system to a certain extent. Musk emphasized in a tweet on August 14 that although FSD may not be able to perfectly navigate the roundabout at the beginning, it can indeed navigate during the roundabout. The Tesla boss predicted that it may take about a year to truly master roundabouts around the world. In fact, previous versions of Autopilot had difficulties passing through roundabouts, and it was difficult to pass smoothly without human intervention. Previously, a Tesla enthusiast released a video of a Model 3 failing to successfully drive out of a roundabout in Autopilot mode, and the driver had to take over the driving task.

  

Tesla's FSD beta test proves that improved understanding is a prerequisite for Tesla Autopilot to work more accurately. Another video posted by James Locke, a Model 3 owner who obtained the FSD beta test version, shows that this Model 3 navigated out of the roundabout relatively easily. What impressed Locke most was that no human intervention was required in this process, and Autopilot completed the driving task independently.

  

Dojo will soon be released with the updated FSD beta, which will prove to be the answer to all of Tesla's previous problems. With this newer, more sophisticated system that is able to take in more information about the terrain, surroundings, and obstacles, Autopilot will be more precise than ever before. With new FSD beta videos being released regularly, people will also see that Autopilot's capabilities are improving every day.


Keywords:Tesla Reference address:Tesla FSD's secret weapon: Dojo neural network

Previous article:MIT uses 19 neurons to achieve autonomous driving control
Next article:ADAS boosts the demand for in-vehicle cameras, and domestic in-vehicle lens manufacturers enter the "fast lane"

Latest Automotive Electronics Articles
Change More Related Popular Components

EEWorld
subscription
account

EEWorld
service
account

Automotive
development
circle

About Us Customer Service Contact Information Datasheet Sitemap LatestNews


Room 1530, 15th Floor, Building B, No.18 Zhongguancun Street, Haidian District, Beijing, Postal Code: 100190 China Telephone: 008610 8235 0740

Copyright © 2005-2024 EEWORLD.com.cn, Inc. All rights reserved 京ICP证060456号 京ICP备10001474号-1 电信业务审批[2006]字第258号函 京公网安备 11010802033920号