Translation:

Main Body:

1399 words in total

Estimated reading time: 6 minutes

Continuing from the previous article:

Tesla AI Day from a Layman’s Perspective (Part 2): Visual Recognition

After completing the visual recognition, the neural network knows exactly what objects are in the image, including the location, speed, and direction of each object. For example, it knows where the road is, the edge of the road, traffic lights, ground markings, and signs. It also knows which key index markers have been encountered recently, which are stored as a sequence of key information snapshots / indexes.

Next, the AP system will do three things.

First, it will determine the intentions of all of these moving objects for their next step. Second, it needs to adjust the original planned driving path based on the next step intentions of these moving objects. Third, it needs to tell the car how to execute the next step.

These seemingly simple steps can become very complex in some situations. Let’s take an example. When the car is driving on the far right lane of a three-lane road and there are several other cars in the middle lane, the route requires a left turn at the intersection ahead. To execute this turn, the car needs to first merge between the two cars in the middle lane, and then switch to the far left lane. All of this needs to be done precisely before reaching the intersection—too fast or too slow would affect the merge or other vehicles’ driving. During this process, sudden acceleration or deceleration cannot occur, as this would affect the riding experience.

Currently, Tesla uses a search algorithm based on C++ programming to achieve route planning. In fact, this kind of search algorithm can simulate thousands of different situations—namely, the mutual relationships and travel path possibilities between Tesla and other vehicles. Of course, this is rough planning. According to Tesla, 2,500 different rough planning simulations can be completed per 1.5 milliseconds. All of these rough plans are based on the dynamic simulation of real objects. If Tesla is driving on a road with fewer cars, the rough planning will also be based on the lanes, barriers, etc.

After rough planning is complete, the car will have a 15-second driving route. Then, based on this rough driving path, the car will develop a more detailed plan to make the entire driving process smoother. For example, the radius of the turn, the lane used for driving, and adjustment of speed.

These are also implemented using a search algorithm based on C++ programming, but with slightly different considerations than rough planning. It combines lateral acceleration and deceleration, lateral movement, collision risks, and execution time into a “total cost.” Then, it constantly iterates—searching for the best path to minimize this total cost.Tesla showcased an example to search for a smooth driving path of about 10 seconds using the search algorithm for approximately 100 iterations.

The overall operating speed of this path planning module, combined with the visual recognition module previously mentioned (Tesla AI Day (Part Two): Visual Recognition), outputs a result every 27 milliseconds, which means it can iteratively recognize and plan 36 times per second. Therefore, to ensure a smooth driving experience without being interrupted by unexpected events or sudden objects, the AP system refreshes its perception of the surrounding road and iteratively plans the driving path 36 times per second.

Of course, the AP system is not all-knowing, and it cannot predict whether the driver in the car ahead is distracted by replying to Twitter while changing lanes. However, it can detect the potential danger of collision with the car ahead faster than you, and adjust its own path planning and driving strategy accordingly.

Currently, Tesla is experimenting with an updated path planning algorithm that mainly uses neural networks, instead of C++ programming like before. Choosing to use a neural network for the planning algorithm in the future may be because this type of algorithm can complete tasks faster and may offer more optimal paths.

What may be incomprehensible to us is that one of the most difficult path planning scenarios is actually a parking lot. Tesla has demonstrated that the neural network planning algorithm it is developing can process similar scenarios more efficiently than the current path planning algorithm.

In the next section, we will talk in detail about Tesla’s neural network training.

(to be continued)The following is an English Markdown text translated from the Chinese Markdown text below, with professional formatting and retaining the HTML tags inside the Markdown. Only corrections and improvements to the original text are provided, without further explanations.

The author of this article, @cosmacelf, is a Reddit technology writer. The original title and link are:

“Layman’s Explanation of Tesla AI Day”

https://www.reddit.com/r/teslamotors/comments/pcgz6d/laymansexplanationofteslaai_day

Layman’s Explanation of Tesla AI Day

Introduction

On August 19th, 2021, Tesla held AI Day, an event designed to showcase their latest developments in artificial intelligence software and hardware. In this post, we will provide a layman’s explanation of what was presented during the event.

The Dojo Training Computer

• Dojo is Tesla’s ultra-high-performance training computer for machine learning and AI applications.
• Dojo includes a custom tensor processing unit (TPU) that can process video data, called D1.
• Dojo can train an AI algorithm to visually recognize 1,000 objects within one hour, much faster than previous hardware.

Tesla Bot

• Tesla announced the development of an anthropomorphic robot.
• The robot is intended to be used to perform repetitive or dangerous tasks, such as attaching bolts or removing parts.
• The robot will use the same AI software as Tesla’s self-driving cars.
• The robot will have advanced sensors and will be “friendly,” according to Elon Musk.

Tesla Vision

• Tesla is developing an advanced vision processing system.
• This system is intended to improve the safety of Tesla’s vehicles and enable them to operate autonomously.
• Tesla is also developing new laser scanners and radars to improve the accuracy and range of the vision system.

Conclusion

Tesla’s AI Day presentation showcased their focus on developing advanced AI software and hardware. Dojo, Tesla Bot, and Tesla Vision demonstrate Tesla’s ambition to not only lead the electric vehicle market but also the AI market.

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.