On June 18, the 2021 China Automotive Forum, themed "New Beginning, New Strategy, New Pattern", opened in Shanghai. LandLeader Technology Co-Founder and CEO Huo Jing delivered a keynote speech.
In Huo Jing's opinion, the intelligence of the city end and **** enjoyment intelligence can make the single-vehicle intelligence greatly reduce the cost and accelerate the commercialization of the landing; secondly, a large number of human behavioral data can contribute great value to the coverage-oriented automatic driving research and development. Finally, it is also hoped that LandLead's imitation learning method can extract the value in the data, establish driving behavior and decision-making models, and realize city-side road-side factory-side **** enjoyment intelligence.
The following speech transcript:
First of all, it's a great honor for a startup tech company like ours to take this valuable opportunity to share with you our thinking based on the research of automated driving, AI, and human traffic behavior.
I'm Jing Huo from Roman Roads LandLeader Technology. Today on behalf of our founder Dr. Xing Zhou, I'd like to share with you the topic of thinking about coverage-oriented intelligent driving development and validation. This is also based on the insights we would like to share with you after three years of research on AI, autonomous driving and human traffic behavior.
Just now Minister Wang also introduced our innovative company, I would like to introduce, LandLeader Technology is a team of startups from the world's top universities, with a wealth of automatic driving and a deep understanding of data, we have an excellent educational background, but also has a wealth of product experience, deep plowing the car and human behavior model algorithms, and is good at mathematical modeling, using machines to imitate humans, learning human logical thinking. that learns to think logically about human beings.
You can see our core technical team, mainly graduated from Stanford, Yale and Carnegie Mellon, Cornell and the University of Science and Technology of China, and all have worked in Tesla, Azure, Amazon, BMW, Samsung, Ford, and deep plowing into the automatic driving, artificial intelligence, and is good at mathematical and physical model building.
What we want to share today is that based on the three years since our founding, and the long time we have been working on the foundation of autonomous driving, artificial intelligence, and on human traffic behavior, we found that there is actually a big challenge to really achieve the realization of autonomous driving. This is manifested in everyone is discussing the more popular single-vehicle intelligence, single-vehicle intelligence is a huge challenge in two aspects: one is the cost, and the second is the large-scale reproducibility. Based on this thinking we proposed the city end, road end and factory end **** enjoy intelligence.
First of all, what kind of problem does LandLeader Technology want to solve? The problem to be solved is very simple, that is, to study how human beings actually drive, and to study the real traffic behavior, including the traffic behavior of cars and people. Like Waymo's long-term road testing, it is hoping to collect data to learn and study human driving behavior.
Why do we do this? We just want to study how humans actually drive and study real traffic behavior, both of cars and people, like Waymo does long-term road tests, actually for what? A lot of data collection, and hopefully through this data to learn and study human driving behavior.
Why are we doing this? We want cars and machines that think like humans to be able to drive autonomously, but what's the reality?
You can see this ppt on the left side of so many yellow circles, this is the study we got from the United States, the yellow circles are inside the phenomenon that shows the car and people in the left turn or straight line collision. Very interesting phenomenon, this inside the front are all rear-end events, rear-end events are all a **** sex, the first car in front of either Waymo, or other self-driving vehicles, the rear of a car rear-end all human-driven vehicles, this means what? All the self-driving suddenly stops somewhere, or somewhere a behavior occurs, and the human-driven vehicle in the back can't react in time and has no idea what you're doing with the car, so it leads to an obvious collision between the human-driven car and the self-driving car. That's what you can see every time an incident happens at Uber, including Waymo, Tesla, there's a reason why a lot of times intelligent self-driving can't really think about things in a manual way like a human can.
Here is a conclusion that our artificial intelligence is not up to the thinking of human intelligence right now, and here I have a simple example. You can see this picture, what does this mean? It shows that the blue color is the time of the self-driving vehicle to make a decision, and the colored one is the time of the decision made by a human being while driving, the decision time in blue indicates that it is the self-driving vehicle in the decision making at least 2 seconds or more, while the human being, because the human brain reacts very quickly, when making a driving decision, it passes in a moment, and that's the difference between a machine and a human being.
The other one is the interaction event, for example, the intersection, two cars, if a person interacts with a person, for example, if you make a gesture, eye contact, are about one second and two seconds, the red one is the self-driving car interaction time must be longer than a person, even up to about 5 seconds.
Here we would like to illustrate that the current artificial intelligence is now not up to human thinking behavior.
This is another case, this red and blue is the human group behavior trajectory of two cities. The blue one we call Blue City, and the red one is called Red City. For example, in Japan, people are generally very compliant with traffic rules, and general behavior is regular, here is the presentation of the blue rules. In red, let's say it's a city in China, many people's traffic habits and behaviors are actually very different, and there are no rules to follow. We can see Cut-in behavior is lane changing, in the red city crowd lane changing behavior, you can find that there is no regularity, are their own change their own, can lead to track line is very dispersed, blue, for example, the Japanese city track line is traceable.
This leads to automatic driving is very difficult: single-vehicle intelligence is facing the current challenge of high cost, just now there are colleagues said single-vehicle intelligence to do automatic driving cost is very high. Now Waymo average a car is 400,000 U.S. dollars equipped with single-vehicle intelligent automatic driving, Waymo in the world more than 600, if more than 1,000 is 400M U.S. dollars cost. The other thing is that even in this car is loaded with sensing, loaded with radar, loaded with millimeter wave, at most can sense 6-8 units, this kind of efficiency is relatively low.
For us, we feel that if autonomous driving is to be realized, we all want to produce low-cost, massively replicable driverless, and in a variety of scenarios, including mines, including unmanned trucks, including minibus logistics. But what is it like in reality? A variety of conveyance, radar, sensing, bring high cost, low efficiency, difficult to realize the commercial scene landing.
We put forward the thinking and our goal is how to large-scale more and more efficient collection of data, so that the machine like human beings with a large number of data to learn the human thinking mode and logic mode, our thinking is more from the road side, the city side, our goal is to use the drone, camera, red and green, etc. to collect data.
What is the idea we are proposing? That is, we hope that the single-vehicle intelligence is getting cheaper and cheaper, more and more reduce the cost, and finally reduced to be able to single-vehicle to make decisions, single-vehicle thinking, and the urban end can be a lot of algorithms with data, with the urban end of the **** enjoyment of the intelligence, equal share of self-driving vehicles to reduce the cost of the more commercial landing.
Just now we said that after a large amount of data collection, is also what we LuLin is doing: using imitation learning algorithms to identify, learn human behavior, strategy. This is what we take in Los Angeles, California, drones and roadside cameras to take, 24 hours a day to monitor the entire road situation, with algorithms and imitation learning constantly looking and thinking about how the car is walking, how people walk, to help the machine to better strengthen the algorithm of the data.
What I said just now is to put forward a concept called CDV concept, is the coverage, we hope that a large number of data collection, data algorithms with the coverage to verify, so that the computer like human thinking. The use of coverage in autonomous driving is not the first thing that we've proposed, and it's not the only thing that we're doing, there's a lot of people in the industry that are doing it. As you can see Foretellix, including LATENT, LATENT has now been acquired by Waymo, at that time Waymo is also interested in their use of the concept of CDV to do self-driving coverage.
What is the probability of coverage proposed? The source of power is frankly two kinds, one is a large number of single-vehicle run to collect data, hoping to win with the quantity, the coverage rate is the quality of the win, with a large amount of data rehearsal algorithms, with imitation learning, thinking like a human being, so that we can collect a variety of abnormal scenarios, closer to the human thinking, so that the safety rate to achieve the purpose of the more enthusiastically said 99.99%.
This is a case we did in Wuxi, Wuxi traffic intersection to take the drone and roadside vision and the city end of the camera collected data to simulate the city and traffic, rehearsed the imitation learning to find a variety of traffic behavior, to help us to carry out a large number of data collection and training, to improve our algorithms.
If we do this, honestly this is not an industry to do is the need for the industry, the government to do, we need drones, cameras to do, including municipal data, Internet *** with the collection of data, a large amount of data collection to the future can be carried out after this algorithm.
Finally, our conclusion is: 1, the urban end of the intelligence, **** enjoyment of intelligence can help our bicycle intelligence greatly reduce the cost and accelerate the commercialization of the landing; 2, a large number of human behavior data can contribute to the coverage-oriented automatic driving research and development of great value; 3, we hope that our LandLeader Technology's imitation learning approach can extract the value of the data to establish driving behavior, decision-making model. Doing this is a relatively huge project, and we are grateful to Auto Shanghai for giving our startup such a great opportunity to fully learn, research, and demonstrate our contribution here.