Author: Lifelong Learning
Introduction
As autonomous driving is gradually becoming mature, various manufacturers have also entered the stage of solving the long-tail problems Corner Case. To achieve true human-like continuous lifelong learning, data-driven solutions are a must. As a leader in the Chinese autonomous driving field, Momenta has been established for 1000 days, always maintaining the lead in AI autonomous driving technology, and has also taken the lead in the data-driven core of autonomous driving 3.0 era.
Data and technology drive autonomus driving into the 3.0 era
In today’s data-driven autonomous driving field, Tesla has already led the world into the autonomous driving 3.0 era, and Momenta is most likely to become the first Chinese company to enter the autonomous driving 3.0 era. Looking back on the development process of autonomous driving, we can divide the past decade’s technological development into three stages: the earliest hardware-driven mode, which we call the 1.0 era of autonomous driving; the software-driven mode of recent years, which we call the 2.0 era of autonomous driving; and the data-driven mode, which is about to happen and will continue to develop, which we call the 3.0 era of autonomous driving. Each era has typical technical features, and the degree of technological development is significantly different due to different driving modes and tools.
In the autonomous driving 3.0 era, data is king. The more data and scenarios one owns, the more advanced algorithms will be iterated, and the autonomous driving experience can be safer and smarter. Currently, the assisted driving mileage generated by Momenta’s assisted driving products has exceeded 17 million kilometers. Among the Chinese autonomous driving technology companies today, this achievement is second to none.In the era of data-driven technology, it is required that data should train itself. In the past, different types of AI models were used for different types of tasks in the field of deep learning, and CNN was the absolute mainstream choice for processing two-dimensional information in the CV field. However, since 2020, the Attention mechanism has made significant breakthroughs in the CV field. It has been noticed that the transformer structure based on the Attention mechanism can become an effective general AI model paradigm. With this, various large-scale models based on the Attention mechanism have emerged, such as Graph Attention and other variations, which can accept various inputs of different modalities, such as language, image, video, and speech, and can output multiple modalities, achieving amazing results in various fields. Two years ago, Momenta initiated the landing development of transformer large-scale models based on the Attention mechanism in the autonomous driving industry. The challenges brought by large-scale models, such as high computing power requirements, high training costs, and high landing difficulty, are being addressed by Momenta through low-carbon supercomputing, improving the design of on-board models and chips, and letting large-scale models play a greater role through data organization.
The era of autonomous driving 3.0 first means a technological leap in perception ability, hardware costs, computing power, cognitive ability, and other aspects compared to the previous era. At the same time, to successfully compete in the era of autonomous driving 3.0, it also tests the forward-looking layout of technology companies. Only by in-depth research and mastery of cutting-edge theories can they gain an advantage in the era of cross-era. Most importantly, the era of autonomous driving 3.0 fully demonstrates the commercial landing capability of technology companies. The rapid landing of cutting-edge theories in commercial engineering requires not only sensitive perception but also the most important engineering capability in autonomous driving. Through Momenta’s growth process, it can be found that Momenta has been preparing for the era of autonomous driving 3.0 and continuously making progress towards it. In terms of perception, cognition, and pattern construction, Momenta has been forward-looking and data-driven. Today, Momenta takes the lead in the era of autonomous driving 3.0, which will undoubtedly bring the era of autonomous driving in China to a new chapter.
Conquer urban scenarios. The data intelligent system MANA is upgraded.In the city of NOH, assisted driving faces various challenges, including frequent maintenance of urban roads, dense large vehicles, narrow lane-changing spaces, and diverse urban environments. To solve these difficult situations, a large amount of scene data is necessary and must be understood like humans in order to maximize their value. MOMAI Intelligence’s MANA is addressing these pain points with its brand new upgrade.
Firstly, MANA uses an unsupervised learning method with large-scale production vehicle data to create model effects. Compared to training with only a small amount of labeled data, the training effect is increased by more than 3 times. This allows MOMAI’s data advantages to be efficiently transformed into model effects that better adapt to the various perception tasks of autonomous driving.
Secondly, MANA treats massive amounts of data equally. Facing the difficulty of “data efficiency” with huge data scale, MANA builds an incremental learning training platform, which extracts part of the existing data and combines it with new data to form a mixed dataset. During training, the new model and the old model are required to output as consistently as possible and fit the new data well. Compared to conventional methods, this saves 80% of overall computing power and increases response speed by 6 times.
Thirdly, MANA has stronger perception ability. By using a time-series transformer model to perform virtual real-time mapping in the BEV space, the output of the perception of lane lines is more accurate and stable, allowing urban navigation autonomous driving to break away from reliance on high-precision maps.
Fourthly, MANA has more accurate perception ability, and there is no longer any signal light of a vehicle that is unidentifiable in China. By upgrading the on-board perception system, MANA specifically recognizes the status of brake lights and turn signals. This makes the driver feel safer and more comfortable when dealing with scenarios such as sudden braking and emergency lane changes.
Fifthly, MANA’s cognitive ability has once again evolved. Faced with the most complex urban scene at intersections, MANA introduces highly valuable real traffic flow scenarios into the simulation system. In cooperation with Alibaba Cloud and the Deqing government, we introduce the most complex urban scene at intersections into the simulation engine, build an autonomous driving scenario library, and through real simulation verification of autonomous driving, the timeliness and micro-traffic flow are more realistic. It effectively solves the problem of city intersections, which was a “hard nut to crack.”
Finally, through this comprehensive upgrade, MANA’s cognitive intelligence is entering a new phase. By deeply understanding a large amount of human driving behavior that covers the whole country, learning common sense and human-like actions, MOMAI’s assisted driving decisions are more like humans, and it can choose the optimal route based on actual situations to ensure safety, giving drivers a more experienced driving feel.With the new upgrade of MANA, Hover has released the “China’s first large-scale self-driving scene library based on the car-road collaborative cloud service”, which is also China’s first self-driving scene library generated from traffic data. This will further accelerate the maturity of China’s self-driving capabilities and the collaborative development of car-road-cloud.
Essentially, MANA’s rapid evolution has significant importance for Hover to win the urban assistant scene battle. As we all know, enterprises with more scenes and data in the auxiliary driving field will have more stable and secure auxiliary driving capabilities. Hover’s intelligent data system, MANA, has accumulated 3 million hours of China’s road driving scene library, has a learning duration of over 310,000 hours, and a virtual driving age of 40,000 years. It can be said that MANA’s powerful and rich scenes and learning ability will bring a qualitative leap to Hover’s intelligent driving capabilities. In the battle of urban assistant scenes, MANA is Hover’s absolute weapon.
NOH is approaching, and Hover is making more understanding of China’s city navigation assistant driving
According to industry experts, 2022 is the “NOH Year”, and many technology companies have delivered NOH solutions and gradually moved from high-speed to urban scenes. The ability of automatic driving to cope with the complex traffic conditions in cities has always been the most concerning issue in the industry. As we all know, currently self-driving or assistant driving heavily relies on high-precision maps, which leads to the problem that if the road environment changes, the offline map data is no longer trustworthy and all algorithms will fail in such scenarios. Based on this background, the hot technical route of heavy perception and light maps emerged, and Hover has developed urban NOH navigation assistant driving based on the MANA system, using the “heavy perception, light maps, and high computing power” technology route.
The algorithm of deep perception effectively solves the problem of high-precision map dependency. First of all, it should be made clear that “deep perception, light map” does not mean that high-precision maps are not needed, but that based on actual road conditions, real-time perception is achieved with the support of large computing power and the MANA intelligent system to construct high-precision maps for vehicles to complete automatic driving operations. According to reports, the MOOV NOH of Haomo City not only includes five bright spot functions such as “intelligent recognition of traffic lights, intelligent left and right turns, intelligent lane changing, intelligent evasion of obstacles-static, intelligent evasion of obstacles-dynamic”, but also launches the “Intelligent Traffic Flow Processing” function. The emphasis on perception by Haomo NOH can not only cope with changing road environments with ease but also greatly enhance the user experience.
As a native Chinese enterprise, Haomo Zhixing is obviously more familiar with the characteristics and standards of Chinese urbanization roads. It has been reported that while Tesla’s FSD has been popular and sought after in the United States, it has experienced constant accidents on Chinese roads. It is precisely the change of environmental scenes that has led to the strong “discomfort with the environment” encountered by Tesla FSD deployed on Chinese roads. The road environment, traffic signs, and driving habits of human drivers are different in different regions and areas, which is why data is so important in the era of automatic driving 3.0. Only by mastering unique and accurate data can algorithms be iterated to their optimal level in that scenario. With more data, Haomo City NOH undoubtedly understands the rules and processing methods of Chinese-style roads better. Haomo CEO Gu Weihao also confidently stated that Haomo City NOH will be a navigation-assisted driving system that understands Chinese urban road conditions better and will soon be mass-produced and delivered. At the Chengdu Auto Show in August, Wei Pai announced that the new Mokka DHT-PHEV laser radar version will be equipped with Haomo City NOH and will be mass-produced in September and delivered by the end of the year.
Summary
2022 is the beginning of the competition in the second half of the autonomous driving industry, which puts higher requirements and expectations on autonomous driving companies. With its leading layout of cutting-edge technology in the industry, Haomo Zhixing has deployed related technologies in the passenger car and last-mile logistics delivery fields. Meanwhile, with its unique understanding of Chinese-style urban scenarios and the new upgrade of its data intelligence system MANA, the starting advantage of Haomo is gradually expanding. In the era of data intelligence technology-driven autonomous driving 3.0, Haomo Zhixing is most likely to become the representative of China’s autonomous driving companies and take the lead in bringing China’s autonomous driving into the 3.0 era.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.