In 2023, a traffic accident involving Cruise, an American autonomous driving company, occurred in San Francisco, USA, once again sparking discussions and concerns about the safety and reliability of autonomous driving.

For many years, autonomous driving has attracted widespread attention due to its potential to revolutionize the paradigm of transportation. However, the question remains: when can autonomous driving be commercially deployed on a large scale?

There is no unified consensus or industry standard on how to test and evaluate the safety of autonomous driving. A research report from the RAND Corporation in the United States shows that autonomous vehicles may only reach an ideal state of safety after being tested for 11 billion miles (about 18 billion kilometers).

Advertisement

This "billion-kilometer" problem, if completed under natural conditions, is full of challenges in terms of environmental complexity and other factors, which has become a bottleneck problem restricting the application and landing of autonomous driving.

Feng Shuo, an assistant professor at Tsinghua University, summarized the scientific issues behind it and proposed a new idea of "continuous spatiotemporal intelligent environment testing". He established the theoretical and methodological system of "equivalent accelerated testing of autonomous driving vehicles", breaking through the limitations of fragmented scene testing, and providing a solution to the low efficiency problem of autonomous driving testing.

Revealing the scientific challenges behind the safety of autonomous driving vehicles for the first time - the disaster of sparsity, he opened up an accelerated testing and sustainable learning framework for safety based on generative artificial intelligence, and improved the speed of simulation and real vehicle testing by 3 to 5 orders of magnitude. Feng Shuo became one of the Chinese candidates for the "35 Innovators Under 35" of the MIT Technology Review in 2023.Providing Theoretical Foundations for Solving the Low Efficiency of Testing

Feng Shuo graduated from the Department of Automation at Tsinghua University with both his undergraduate and doctoral degrees, under the guidance of Professor Zhang Yi. In the following five years, he studied and worked in the research group of Professor Liu Xianghong (Henry X. Liu) at the University of Michigan as a joint-trained Ph.D., postdoctoral fellow, and research assistant.

As his research during his doctoral studies progressed, Feng Shuo, who was originally engaged in the research direction of multi-intelligence vehicle control, found that the large-scale application of autonomous driving vehicles faces a huge challenge. Before solving this problem, most of the research on multi-intelligence vehicles based on the large-scale application of autonomous driving was mostly on paper and difficult to make substantial progress.

Therefore, in his fourth year of doctoral studies, he resolutely changed his academic direction and began to focus on the research of safety testing issues for autonomous driving vehicles.

Moreover, following a research method from simple to complex and gradually deepening, he successively focused on the generation of testing environments at the "single point level," "section level," and "network level."To address the challenges brought by high-dimensional problems, he further formulated the scene generation issue as a Markov decision process problem and developed a reinforcement learning algorithm that learns critical behaviors by evaluating their scenario criticality.

For the single-point car interaction driving environment, a theory of criticality sampling based on importance was proposed, providing a theoretical foundation for solving the problem of low test efficiency. Based on this, a method for generating a single-point car test environment was proposed, significantly accelerating the single-point car testing process [1-3].

The related achievements won the 2020 IEEE Intelligent Transportation Systems Society Best Ph.D. Dissertation Award (an authoritative Ph.D. dissertation award in the field of intelligent transportation, only three papers are selected globally each year).

"I, together with my research group, started from scratch to explore the generation of autonomous driving test scenarios, in order to test the safety of autonomous driving better and faster. In this research stage, we found mathematical tools to solve the safety of autonomous driving and preliminarily verified them," said Feng Shuo.

It should be understood that the testing of autonomous driving is divided into two stages. The first stage is the fragmented scene, which is mainly threshold testing. The second stage is the spatiotemporal continuous traffic flow test, to explore the safety performance of autonomous driving under complex interaction scenarios.During his postdoctoral period, he realized that as the capabilities of autonomous driving become more advanced, fragmented scenarios can no longer meet the safety assessment requirements of autonomous driving. Therefore, Feng Shuo began to study the theoretical extension to more complex scenarios, focusing on the multi-vehicle interactive driving environment at the road segment level.

Based on this, he, as the sole first author, proposed the "sparse adversarial sampling" theory in Nature Communications, which means that the natural-adversarial driving environment can continuously generate test scenarios for testing autonomous driving in any driving environment, solving the "curse of dimensionality" of importance sampling theory.

On this basis, for the first time, he proposed a method for generating multi-vehicle test environments on road segments, significantly enhancing the ability to test the safety of vehicles in continuous spatiotemporal complex environments [4].

Specifically, because the basic input for scenario generation comes from natural driving data, the test mileage on the test track can be approximately converted to equivalent mileage in the same natural driving environment on the road.

Due to the significant increase in challenging scenarios, the exposure rate of safety-critical cases is increased, and testing 1 mile on the test track may be equivalent to driving hundreds or even thousands of miles on public roads.The relevant paper was selected as a featured article in the "Artificial Intelligence and Machine Learning" special issue of Nature Communications, and received the "2021 Intelligent Transportation Systems Best Paper Award" from the Institute for Operations Research and the Management Sciences.

This is the first paper published in the main journal Nature in the field of autonomous driving safety.

In 2022, Feng Shuo returned to his alma mater, Tsinghua University, as an assistant professor and doctoral supervisor, with a research direction in intelligent system testing and verification.

He proposed the theory of dense reinforcement learning for the multi-vehicle interactive driving environment at the road network level, providing a solution to the "Curse of Rarity" faced by artificial intelligence technology in the optimization of low-probability events in high-dimensional space.

On this basis, he proposed a method for generating multi-vehicle test environments in the road network, achieving intelligent generation of test environments, and through the "AI Against AI" method [5] between the intelligent environment and autonomous driving.This method significantly enhances the safety testing capabilities of autonomous vehicles on a large spatiotemporal scale, increasing the testing process by 3 to 5 orders of magnitude (1,000 to 100,000 times). It is expected to greatly reduce the cost of autonomous driving testing and research and development, and accelerate the speed of its application.

The autonomous driving company Google/Waymo commented on the study as "accelerating the system verification process" and "significantly reducing the required testing mileage."

In this study, through an intensive deep reinforcement learning method, it allows the neural network to learn from critical safety events in dense information and achieve tasks that are difficult to complete with traditional deep reinforcement learning methods. The superiority of this method has been theoretically proven and tested on a high-speed highway testing track with highly automated vehicles.

In March 2023, the related paper was published as a cover article in Nature with the title "Dense Reinforcement Learning for Safety Validation of Autonomous Vehicles."

Feng Shuo is the first author, and Professor Liu Xianghong from the University of Michigan is the corresponding author. It is known that this is the first paper published in the main journal of Nature in the field of autonomous driving safety.The new theory and method of "AI Against AI" is expected to form a new paradigm for testing and developing the next generation of machine intelligence, exerting a huge influence on the large-scale application of safety-critical systems, including autonomous driving, aerospace, medical machinery, intelligent nuclear power, and smart grids.

Feng Shuo said that, based on this research, using AI to help AI in the future may form a process of mutual learning. One AI is the coach, and the other AI is the student, requiring the two to promote each other to improve their abilities.

Accidents in real traffic environments are very rare, and the human accident rate is generally around 10-6 or even lower. How to model this long-tail low-probability event with high precision has always been a difficult point in the field. Feng Shuo and his collaborators used generative methods to produce high-precision natural driving environments, making the intelligent testing environment more similar to reality [6].

Feng Shuo said: "Through a series of studies, we found that the safety testing and safety training of autonomous driving are like two sides of the same coin. If you do not know the safety and quantified performance of the algorithm, you do not know whether the algorithm has been improved, and you do not know how to improve safety."

Will actively explore the direction of high-value data generation.Feng Shuo stated that embarking on the path of scientific research is inseparable from the guidance of the school, the teachings and examples of his two mentors, and the support of his family.

He said: "Professor Zhang Yi encouraged me to be a jointly trained Ph.D. student, helping me to explore a completely new doctoral topic; after I returned to the country to teach, he supported me in establishing my own laboratory. Professor Liu Xianghong, on the other hand, 'hand by hand' led me into the field of autonomous driving."

What impressed Feng Shuo the most was the "fully immersive" mode of writing papers. He once spent 3-5 consecutive days in the office, face to face with Professor Liu, writing and revising papers. Through immersive thinking and discussion, they enhanced their understanding and comprehension of the issues. This method was highly efficient and has become a phased work mode for Feng Shuo.

At present, the commercialization of autonomous vehicles has "slowly" entered the L3 era, yet L3 cannot fundamentally change the traffic pattern. Feng Shuo believes that the current stage is still about 1 to 2 orders of magnitude away from the large-scale commercialization of autonomous driving with safety performance reaching L4 or even L5.

He expressed that the field of autonomous driving has not reached the expected development progress, and the fundamental reason is the existence of key scientific problems that have not yet been solved. Breaking through these problems is not enough with the power of the industry alone; it requires the joint efforts of the academic community.On the other hand, the development of the industry also requires supervision. Safety is a major issue involving human society. With the rapid development of AI technology, the importance of AI security issues is becoming increasingly prominent.

"As a scientific researcher in this industry, we should strive to be ahead of the industry to perceive these potential industry bottlenecks, lay out academic research in advance, and clear the obstacles for industrial development," he said.

At present, Feng Shuo is promoting the industrialization of related technologies. Relying on traffic big data and large model technology, a large model of traffic behavior oriented to the safety testing and training of autonomous driving is constructed to achieve high-fidelity, strong interaction, and large spatiotemporal scale of traffic behavior simulation.

The scenes he focuses on include: accelerating the safety simulation testing of autonomous driving vehicles by generating intelligent test traffic environments; accelerating the training efficiency of autonomous driving by generating high-value training data and dense learning algorithms.

At this stage, Tesla collects autonomous driving data through actual collection methods, but it needs to be understood that only one in ten thousand is valuable data. Recently, building end-to-end capabilities has become a widely discussed topic."Without the generation of high-value data, it is difficult to get the entire system of autonomous driving to run, so this is also the direction we will focus on in the future, hoping to accelerate the arrival of the L4/L5 era of autonomous driving."

When it comes to the innovative elements in research, Feng Shuo believes that one should not be overly attached to the "low-hanging fruit." Research is like treasure hunting in the deep mountains. After achieving certain breakthroughs, one should not rush to "set up camp," but instead have the courage to continue on the road and venture further into the "uncharted territory," striving to explore the ultimate treasure.