How does Tesla and Waymo handle the most important item "data" in automatic driving car realization

In order to realize an automatic driving car that will be developed for the realization of the 2020's, in order to analyze and extract useful contents from a huge amount of data, artificial intelligence (AI) which judges on behalf of human beings We need to make many things learn. Mr. Eulon Mask serves as CEO who is said to be at the forefront of realizing automated driving carTeslaAnd that is a brother company of GoogleWaymoAlthough there are two companies, there is a big difference in the approach of development by both companies.

Tesla vs. Waymo: who's winning the race for self-driving cars - The Verge
https://www.theverge.com/transportation/2018/4/19/17204044/tesla-waymo-self-driving-car-data-simulation

Human drivers can predict and understand beforehand how the same human driver will behave. On the other hand, however, in the case of an automatic driving car where the computer makes decisions, it will be acquired by learning based on a large amount of data. Waymo and Tesla of Google are said to be at the forefront of technology development.

Both companies have accumulated data every day to enable automatic operation, but the approach is completely different. Tesla has already sent hundreds of thousands of commercial vehicles to the world, and Tesla acquires travel data in the real world every day through computers installed in each vehicle. Meanwhile, Waymo learns artificial intelligence (AI) by doing an enormous number of simulations in an environment where the real world is reproduced mainly by a powerful computer.


It is said that popularization of automatic driving cars can reduce the number of people who die in traffic accidents, but the merit is not limited to that alone. It is said that there are factors that motivate the spreading of monetary terms by realizing an automatic driving car. Intel believes that an automobile will produce $ 800 billion annually (about 82 trillion yen) annually in 2030 and will expand to $ 7 trillion (about 750 trillion yen) per year in 2050 I am predicting.

Tesla is promoting the development of an automatic driving car using a lot of data acquired from the car the customer is running. It is difficult to grasp the total number of mileage data actually acquired by Tesla, but according to the remarks of a person led by the automatic driving car division at the time of 2016, Tesla estimated 780 million miles (about 13 It is said that 100 million miles (about 160 million km) are the data when driving by the autopilot function. In addition, Mr. Mask revealed that Mr. Mask acquired 3 million mile (about 4.8 million km) of traveling data per day in the summer of the same year, and in July 2017 the total of the running data You know that the distance has reached 5 billion miles (about 8 billion km)I had made it clear.

Bymariordo 59

Most of the log data is not for autopilot running, but Tesla's vehicles collect information on the surroundings and upload them to Tesla's server even during manual operation. With this function called "shadow mode", Tesla is now able to take advantage of real-world running data equivalent to hundreds of millions of kilometers.

It has data comparable to Tesla that in the beginning of 2018 the total distance traveled by simulation reached 5 billion miles (about 8 billion km)Publish on TwitterIt was only Waymo who did it. Waymo also has acquired 5 million mile (about 8 million km) of automatic driving data on public roadsI am clarifying. The actual mileage that Tesla and Waymo has acquired so far are larger than the sum of all the achievements of other companies developing similar automatic driving cars.

Automatic driving data acquired by Waymo is restricted to those acquired from experimental vehicles, which are said to be owned by the company from 500 to 600, and Tesla, which already runs around 300,000 vehicles around the world Compared to, it is small scale. Also, Waymo's vehicles are allowed to run in Texas, California, Michigan, Arizona, and so on. On the other hand, Tesla is acquiring a huge amount of data compared to Waymo, but its data is in the "semi-automatic operation" state, so it has the disadvantage that it is not data for pure automatic driving function . Waymo plans to acquire even more data by increasing the actual traveling vehicle significantly in the future, but until then until then, there are circumstances in which it is inevitable that data acquisition by computer simulation will be main.


There is also a difference in the type of data that is acquired between the two companies. Waymo has three radar sensors and eight cameras, in addition to installing three LIDARs on each vehicle to grasp the distance to the object by irradiating countless laser beams. One Tesla also has 8 cameras, 12 ultrasonic sensors, and a front camera, but LIDAR is not installed. Regarding LIDAR, we admit that the mask CEO is also very important, but LIDAR is still expensive and bigger in size and has problems to be solved with durability such as adoption for products for general consumers It is still time to take some time to do. However, if Tesla realized automatic operation without LIDAR, it can be said that it can be a big advantage of Tesla.


What matters as much as the data to be acquired is how to process the data and reflect it on learning. It can be said that Waymo has an advantage in this respect. In Waymo, it is possible to update the feedback loop at a high frequency by analyzing the data obtained on the simulation, learning AI, and performing further simulation in that state. Therefore, it is possible to repeat learning with efficiency superior to running in real world.

Processor manufacturer NVIDIA is showing the presence in the simulation world. In March 2018, a simulation system that provides a simulation environment in the virtual worldDRIVE ConstellationWe are now offering possibilities for each company to develop using this technology in the future.

Also, the advantage of development using a simulator is that it is easy to increase learning efficiency. It should be able to be felt if you are actually driving a car, but most of the driving scenes of the car can be said to be "idle running time" in which nothing happens. In the simulation, it is possible to deepen learning more efficiently than running in real society by eliminating this idle running time and intensively reproducing important points and learning it.

Nonetheless, the simulator is not all-purpose, so it's impossible to escape because the science is a virtual space that reproduces the real world. Therefore, in order to realize an automatic driving car, it is still indispensable to acquire driving data in real society. Whether the automatic driving car is really safe or not, it seems to be said that it will be obvious for the first time since the first time the automatic driving car runs through the town.