The International Conference on Computer Vision and Pattern Recognition (IEEE/CVF Conference on Computer Vision and Pattern Recognition, hereinafter referred to as "CVPR"), as one of the top academic conferences in the field of computer vision and even AI, is used by the autonomous driving industry and academia as a window to observe the global technology development direction every year. For example, in 2021, it disclosed its FSD technical ideas and latest progress, launched occupancy networks in 2022, and proposed a generative world model in 2023...... Each release sparks a heated debate in the industry about the concept of autonomous driving technology.
The just-concluded 2024 CVPR has set a new high in a decade with nearly 20,000 papers submitted. The 10 award-winning papers provide new ideas for the current autonomous driving from different perspectives such as simulating the oscillation of natural objects, from perception recognition to simulation and testing.
In addition, unlike technical evangelism and dissemination through papers alone, some enterprises and research institutions have begun to actively guide and set research topics by holding challenges and other methods, so as to gather scientific research forces more efficiently for academic research.
This year, the CVPR Autonomous Driving International Challenge focuses on cutting-edge issues, with a total of 7 tracks, including the application of large language models in autonomous driving, end-to-end autonomous driving, and mapless driving, attracting nearly 500 teams from 28 countries and regions around the world. The organizers and participants of the event are characterized by a mix of industry and academia, including academic research institutions such as Shanghai Artificial Intelligence Laboratory, Tsinghua University, and University of Tübingen, as well as technology companies such as Meituan, Wayve, and NVIDIA.
"Potential stocks" in the field of autonomous driving
Different from Tesla, Huawei, Waymo, Baidu Apollo, Pony.ai and other CVPR intelligent driving stage in the center of the "traffic responsibility", Meituan and other companies focus on the field of autonomous delivery vehicles, which is more low-key, and it is also the first time to sponsor CVPR events.
The automatic delivery vehicle mainly provides instant delivery services within three to five kilometers for citizens and enterprises. How to use technical means to improve the efficiency of goods distribution and reduce the cost of distribution is one of the important indicators of automatic delivery vehicle technology research.
As an external consequence, autonomous delivery vehicle companies are receiving much less attention than other types of autonomous vehicle companies, and they do not match their market size and growth potential.
However, in fact, autonomous delivery vehicles are facing a market of nearly 100 billion orders per year.
According to the "2023 White Paper on China's Instant Delivery Industry Trends" released by Frost & Sullivan, the order scale of China's instant delivery industry will reach about 40.88 billion orders in 2023, a year-on-year increase of 22.8%, and it is expected that by 2028, the market size will reach 81.31 billion orders, maintaining an average annual double-digit growth of 14.7% in the next five years; At the same time, according to the data of the State Post Bureau, in 2023, the postal industry's mail and delivery business volume will increase by 16.8% year-on-year, which is also in the high-speed growth range.
At the same time, the unstructured road environment faced by autonomous delivery vehicles also provides rare scenario data for the research and development of autonomous driving technology. This is an area that is more exciting for technologists in terms of the complexity of the scene.
An algorithm engineer who has just entered the autonomous delivery vehicle industry said that in the past month or two, more rare scenarios have been encountered than in the previous two years. During the busy farming season, corn cobs cover the road; Wrong-way bicycles, electric scooters, and even wheelchairs that appear from time to time; small children laughing and playing on non-motorized lanes and sidewalks; A drunken passer-by lying on the ground ...... "Every day the algorithm research and development has the feeling of opening a blind box, and the automatic delivery vehicle often contributes hidden money." The above-mentioned engineer said that how to efficiently use rare scenes for technology research and development and scene landing to speed up is one of the biggest problems faced by automatic delivery vehicles.
In addition, under the existing traffic rules, the left turn of an autonomous delivery vehicle is also much more complicated than that of other motor vehicles, and it is necessary to wait for two green lights in different directions like a pedestrian to make a "second aisle". When the green light is on, pedestrians on both sides of the road cross the road like dumplings, and the automatic delivery vehicle must not only ensure its own traffic efficiency, but also not hinder the passage of other pedestrians and non-motorized vehicles.
Large model + end-to-end, providing ascending dimension solution
Although autonomous delivery vehicles are often lucky enough to open "hidden blind boxes" in technical scenarios, in terms of the total amount, such scenarios still have the characteristics of low probability of occurrence, many types and unpredictable in advance.
"The low probability of occurrence makes it difficult to accumulate data to solve problems in a data-driven way, and the variety makes it difficult for us to exhaust the problem by adding manpower. The traditional way of dividing autonomous driving by modules is no longer enough to break through the existing bottlenecks. Yun Feng (pseudonym), an algorithm engineer at Meituan's autonomous vehicle delivery department, said that in order to meet this challenge, Meituan began to study and explore the use of large models and end-to-end solutions to solve problems.
Yun Feng said that the first step is to generate scarce scene data through large model technology, and then combine the rare data accumulation of automatic delivery vehicles for several years to simulate various problem scenarios and simulate various sensor data to enhance the diversity and coverage of training datasets and improve the efficiency and effect of model training. The next step is to realize end-to-end autonomous driving through a segmented and guided path. Finally, through a large amount of data, the autonomous driving model can learn the "knowledge" precipitated in the large model to achieve more intelligent autonomous driving.
At present, Meituan's autonomous delivery vehicles have applied multi-modal large models to the identification of complex scenarios such as temporary construction road conditions, allowing vehicles to detour more intelligently and significantly improve the traffic efficiency of similar complex scenarios.
However, technological advances in the field of autonomous driving have not been smooth sailing. In the process of iterating technology, the so-called "seesaw effect" is often encountered, that is, the technical performance of one aspect is improved, and the technical performance of another aspect is deteriorated.
For example, the improvement of accuracy may lead to a decrease in the recall rate, the improvement of the effect of general scenarios may lead to a decrease in the effect of specific scenarios, and the improvement of safety capabilities may affect the efficiency of traffic. Continuous technology development and practical testing are indispensable for this purpose.
Therefore, through the continuous actual road testing of the vehicle and the feedback of real data, the model can be continuously optimized, the algorithm can be improved, and new problems in the application of autonomous driving can be solved.
Integration of production, education and research, press the acceleration button on the ground
As an academic conference, CVPR has attracted wide attention from autonomous driving companies because as a "strong application-oriented" and "strong technology-driven" industry, it is difficult for autonomous driving to rely on a single enterprise and a single research institution to quickly achieve large-scale promotion and scenario implementation of technology, and it requires multi-field, multi-disciplinary, and even cross-industry cooperation.
With the expansion of the scale and scope of services of related enterprises, the iteration of autonomous driving technology has also entered the deep water area, which requires the deep integration of various forces.
On the one hand, relevant enterprises have real scene data and operational experience in practice, and on the other hand, research institutions in different academic fields such as mathematics, computer science, traffic engineering, and materials science also need a real verification environment for how to quickly implement their cutting-edge achievements in industrial practice.
Therefore, Meituan and other enterprises can not only accelerate the large-scale implementation of the industry, but also provide forward-looking guidance for the future development of autonomous driving through the in-depth competition of top technical teams.
Taking CVPR as an example, there are many examples of industry and academia working together to overcome technical challenges. In 2023, Shanghai Artificial Intelligence Lab cooperated with SenseTime and others to propose the industry's first autonomous driving model UniAD with integrated perception and decision-making, and finally won the CVPR Best Paper Award that year.
This year, more than one of the 10 CVPR award-winning papers was submitted in the form of a joint submission between academic research institutions and enterprises.
The automatic delivery vehicle is deeply integrated into the urban transportation operation network and has a high degree of interaction with the daily life of citizens. At present, in addition to co-organizing short-term competitions with academic and research institutions, Meituan Autonomous Delivery Vehicles has also cooperated with Tsinghua University in terms of course teaching, student practice, and topic exchange through the establishment of the "Tsinghua University-Meituan Digital Life Joint Research Institute" with Tsinghua University, so as to provide a real practice environment for talent training in the frontier field of autonomous driving.
epilogue
With the development of the economy and the improvement of the consumption level of urban residents, consumers' demand for products has changed from less variety, large quantity and low frequency to more variety, small batch and high frequency, and instant retail has developed rapidly. This has also brought about a change in the demand for distribution, that is, many orders appear in convenience stores, fresh supermarkets and other places just a few kilometers away from the public.
If the speed of goods flow is taken as a reference index, the 3 to 5 km distribution link is the most complex link in the entire social logistics chain. Some data show that the cost and time investment in the distribution link accounts for more than 30% of the whole distribution process. At present, this type of delivery is mainly delivered by riders with the help of electric vehicles and legs.
Against this backdrop, emerging technologies such as autonomous delivery vehicles are beginning to scale up and work with riders to alleviate the hardships of riders.
According to incomplete statistics, as of August 2023, a total of 54 regions in China have issued detailed rules for the management of road tests of intelligent networked vehicles, proposed to build 17 national-level intelligent networked vehicle test demonstration zones, opened more than 15,000 kilometers of test roads, issued more than 2,800 test licenses, and the total mileage of road tests has reached more than 70 million kilometers.
Beijing Shunyi and other pilot companies have allowed pilot companies such as Meituan to carry out pilot projects of automatic delivery vehicles no higher than 45km/h on motorways in some road sections, and explore the construction of a "policy test field" for the national "commercial operation stage management system for autonomous vehicles".
Driven by policy encouragement, industrial demand and technology iteration, autonomous delivery vehicles are releasing their potential and moving to the center of the stage.