Talking about the next breakthrough of the development of unmanned driving

A driverless car is a smart car that senses the road environment through an in-vehicle sensing system, automatically plans driving routes, and controls the vehicle to reach a predetermined target.

This article refers to the address: http://

It uses the on-board sensor to sense the surrounding environment of the vehicle, and controls the steering and speed of the vehicle based on the road, vehicle position and obstacle information obtained by the perception, so that the vehicle can travel safely and reliably on the road. It integrates many technologies such as automatic control, architecture, artificial intelligence and visual computing. It is a product of high development of computer science, pattern recognition and intelligent control technology. It is also an important indicator to measure the scientific research strength and industrial level of a country. The national economy has broad application prospects.

In the field of driverless vehicles, every step of the development of technology must be measured by the protection of personal safety. In addition to its development, it will bring people's joy and encouragement, but also raise concerns about its safety.

Google, Trass, Zoox... and more companies are trying to make driving distances for driverless cars as fast as a billion miles. At present, there are two bottlenecks in the development of driverless vehicles: First, the official minimum vehicle requirements. Second, the experimental and test models are still to be optimized.

Engineers based on big data to improve the exercise model of unmanned vehicles, experimental scenes from sunlight to sensors, and then from different angles of sensors to flight obstacles in front of the car and abnormal external behavior.

The problem is that it is not easy or safe to reproduce realistic scenes.

A report from Rand found that if you want to test whether the safety of driverless driving is acceptable, you need proof of tens of thousands of miles or even billions of miles of experimental mileage.

“Even with the most reasonable plan, the existing driverless car will take decades or even hundreds of years to complete the scheduled mileage test. If the test is put on the road, it will be impossible. Mission."

What methods can be used to continuously improve the reliability of unmanned vehicles?

Companies like Uber, Lyft, and Zoox were born in big cities and operate under certain conditions to lower their technical barriers. But this may apply to Ubers around the world, and traditional OEMs choose to bridge the gap with shared technology by constantly updating the car's automation capabilities.

Therefore, we can bypass the traditional technical methods that currently require a large amount of data, and construct a model that can reason and learn small amounts of data. Gary Marcus, who was acquired by Uber last year, spent several years studying the issue, but such learning models have not yet become a reality in drones.

Don't forget the simulation

Simulations from software to hardware are reasonably modeled, providing the possibility for companies to experiment and test their car models.

This includes a variety of application scenarios, including traffic, driver behavior, weather, and road conditions.

Also consider the use of the sensor. How many cameras and radars do you need? Where should they be placed? Which model hardware should I use?

At the same time, flexible randomization is also very important. Based on this, there is no need to tie the team and the reliable driver together on the road.

We have not yet reached the end

Today, products such as Vires, TaSS PreScan, CarSim, Oktal ScanNer and ROS Gazebo offer engineers the possibility to simulate sensors and their mechanisms and mechanical structures. Despite their strengths, they also overlook areas that are critical to simulation, including oversimplifying existing sensor outputs and understanding how the environment affects the complexity of autonomous models.

However, as technology continues to evolve, we must consider how to implement, insert, and test the fusion of hardware and software in a high-fidelity manner.

High fidelity simulation environment

Although it is difficult to simulate the perception of the outside world by most sensors, simple simulations are becoming more common in vehicles.

Optical cameras have been expected due to the unfulfilled promise of low-cost LiDARs and the shortage of high-end units that make the scalability of OEMs and Tier 1 difficult.

There is no error in the analog camera's analog data and the input data, so in order to properly test the perception of the outside world, engineers need to construct a realistic simulation environment. But building a complex analog aperture is expensive and difficult, so no one can simulate the environment in order to build an unmanned car.

About a year ago, I met Craig. At the time he was releasing something he called DeepDrive. I later learned that as one of the early engineers, he used the game to sell $137 million in development costs to reproduce real-world scenarios and support driverless cars by showing high fidelity...

A few months later, Craig joined a small startup called Uber, focusing on research simulation.

A research team at Princeton University detailed the advantages of using GTA V. It will be divided into 100 square miles, 4 million people, 262 vehicles, 1,167 different creatures, 14 weather conditions, and more than 70,000 dynamic sections in urban, rural and woodland environments.

Is it really useful to simulate mileage?

Different views on simulating mileage utility

The support side believes that simulations can be used to simulate rare cases and baseline data, and rare cases are scenarios that are difficult to reproduce or are sufficiently random. If driverless can provide 99% reliability, most scenes have been optimized through simulation. Some future technical iterations of AI or ML allow us to react to extreme situations without prior data preparation.

In addition to special cases, simulation is also very useful for building basic data sets, and further testing is continued on this basis.

The opposing party believes that the corresponding environment is that the simulation environment is not good enough to generate the model efficiently. Often, this is a scenario where the environment interacts with the vehicle and is difficult to reproduce in a real-life scenario. In addition, there are scenes where the image fidelity is too low.

From simulation virtual to reinforcement learning in real life situations

To help solve some of the problems with data quality, researchers are testing the possibility of translating virtual image inputs into real-world models to improve simulation experiments.

Google has released news that although many government agencies are reluctant to use simulated miles as part of the required mileage for autonomous driving tests, this situation may change as the regulation of simulation becomes more explicit.

Simulation is necessary

If the accuracy is high enough, the simulation is valuable. True, the simulation may not solve the last 1% of autonomous driving problems. But if the technology is reliable, then in the future you can let the model accomplish better scene recognition or cope with a wider range of scenarios.

Many companies agree with this. Including Tesla, Zoox, Comma.ai, Drive.ai and Aurora Innovations are actively recruiting simulation engineers.

Outside the driverless field

The use of analog technology can be extended beyond the driverless field. Although we can understand how drones perceive the world around us, we can better understand the potential logic of traffic, driving behavior and even pedestrian behavior.

To take a step back, there are enough specific models and dynamic lives in a simulated environment, so we can better understand the robots, which will interact with our real world and the digital world.

Companies like Improbable have targeted this potential market. Investors have also recognized the hidden value of this technology as an architect of the future simulation world.

We have just touched on the indication of this technology. Many companies are developing this technology, and some startups have begun to develop independent software. As the research progresses, it is expected that a variety of new players will enter the market. Those who are the first to succeed have the opportunity to become early leaders, or to lead others to better stage development.