Every user is a guinea pig. It turns out that this is how Tesla’s autonomous driving data is accumulated.

Text/Zhang Yi

The Rand Think Tank has made an assessment of self-driving cars, that is, for a truly self-driving system to meet the conditions for mass production application, it needs to go through at least 11 billion miles (approximately 17-18 billion kilometers) of road verification.

Based on this estimate, even if a team has 100 self-driving vehicles conducting road tests 24 hours a day, the time required to accumulate data will be measured in "hundred years"...

According to our country’s five-year plan, it may take at least 20 five-year plans to achieve the small goal of autonomous driving. Is it really that difficult? Neither.

According to data recently released by Tesla’s AI department, Autopilot (autonomous assisted driving, hereinafter referred to as AP) has been driven more than 3 billion miles since it was launched in 2015 as of this month. , in which the NOA (Navigation Assisted Driving) enabled driving mileage has exceeded 1 billion miles, the automatic lane change function has been executed more than 200,000 times, and the smart summon has been used more than 1.2 million times.

However, it should be noted that this data is not accumulated by Tesla’s own internal test vehicles. After all, in Tesla’s autonomous driving report submitted to the California Department of Motor Vehicles (DMV) last year, they My self-driving test mileage in California is only 12.2 miles (approximately 19.6 kilometers)...

The 3 billion miles of AP driving data are generated by all Tesla owners who use the AP function. In layman's terms, every Tesla owner who uses the AP function becomes Tesla's guinea pig for free.

The Tesla Ap system has a "shadow mode" that can run in the background even when the Ap function is not turned on and collect real-time operating data under real road conditions to feed back to Tesla Pull the neural network for autonomous driving to learn.

In fact, if calculated according to this statement, then the driving data even when the AP function is not turned on is still useful to Tesla. As early as November 2018, when Tesla had just completed the delivery of its 500,000th vehicle, the total mileage of Tesla vehicles around the world had reached 10 billion miles, almost approaching 11 billion miles of autonomous driving. The conditions for mass production application of driving systems are probably already exceeded.

For comparison, Waymo’s current test mileage on open roads has just exceeded 20 million miles, while the domestic Baidu Apollo self-driving car’s cumulative road test mileage is less than 2 million miles. The gap is almost exponential. , it can be said that the road test data currently held by Tesla is almost unmatched by anyone.

So in road testing, are other self-driving companies represented by Waymo no longer able to compete with Tesla? Of course not.

Actually, there are two types of road tests for self-driving cars. One is real road testing in a development environment, and the other is on a self-driving simulation test platform. Conduct virtual testing. At present, almost all car companies and self-driving technology companies mainly use virtual testing, supplemented by real road testing. There is only one exception, Tesla.

Of course, the reason Tesla has such a huge data accumulation is because it dares to let car owners be "guinea pigs" instead of doing the raw data accumulation themselves. Other players do not have this condition, so they can only use the autonomous driving simulation test platform for testing.

You can think of the autonomous driving simulation test platform as a large-scale game similar to GTA. Of course, the actual simulation test platform is much more complicated than the game. However, the benefit of the virtual simulation test platform is not only time saving, but also that it is controllable. This controllability includes the ability to model various situations that do not often occur in the real environment on the virtual simulation test platform.

For example, extreme rain and snow weather, various nighttime road conditions, etc., and are not subject to real conditions. For example, on a simulation test platform, tens of thousands of self-driving cars can be run at the same time. The daily driving mileage can be measured in millions of kilometers.

Alibaba recently released the world’s first autonomous driving “hybrid simulation test platform”. The virtual test mileage can exceed 8 million kilometers every day, and the time to generate an extreme simulation scenario is only 30 seconds. Every day The number of supported scene constructions reaches millions, which is difficult to achieve in real road tests.

Waymo has also claimed that there are 2.5 scenarios in their autonomous driving virtual test platform every day. Thousands of test vehicles shuttle through it, and each test model learns something new from its experience and shares it with other cars to avoid similar dangers.

Currently, although Waymo’s test mileage on open roads has just exceeded 20 million miles, its simulation testing has exceeded 10 billion miles. Of course, what is unmatched by Tesla is that its 3 billion miles of driving data all come from real road conditions, not computer simulations.

In fact, for the world's major car companies and self-driving science and technology companies, virtual simulation testing mainly relies on means. 90% of the scenarios in self-driving testing are completed through simulation platforms, and the rest 10% is completed through test fields and actual road tests.

After all, the cost and cycle of actual road testing are too high and too long. Extreme weather may occur once every few months, but the virtual test platform can simulate it countless times a day, and not everyone can Have the "courage" of Tesla.