Zero One Automobile is a new energy intelligent heavy truck technology company, committed to creating excellent new energy vehicles and high-level autonomous driving systems through the ultimate new energy efficiency, the ultimate big data application, and the ultimate automation realization, and finally grows into the world's leading transportation robot company.
"The development of autonomous driving involves three core elements: software, vehicles and operating environment." On September 13, 2024, at the 4th Commercial Vehicle Autonomous Driving Conference of Gasgoo, Huang Zehua, founder & CEO of Zero One Automobile, said when talking about how to achieve mass production of unmanned driving. In terms of software, he believes that large models are the best path to achieve mass production of unmanned driving, and the safety problems of large models that are of general concern to the industry can be subdivided and solved from four dimensions: principles, new algorithm technologies, processes, and systems.
Huang Zehua | Founder & CEO of Zero One Automobile
The following is a summary of the speech:
Introduction to Zero One Automobile
Zero One Auto is a new energy intelligent heavy truck technology company, pursuing the ultimate product, ultimate technology and ultimate efficiency, and is committed to building an innovative ecological alliance through cutting-edge data, products and technologies to create excellent new energy vehicles and high-level autonomous driving systems. ZeroOne Auto's vision is to become the world's leading transportation robot company.
In the field of vehicle development, we have already made remarkable achievements: in April this year, the first two models were delivered to end customers in batches, and to date, we have placed orders for nearly 2,000 units, and have successfully delivered about 200 units. Our models are known for the industry's lowest system energy consumption, thanks to the core technologies such as the Rectangle intelligent integrated electric drive system and the five-in-one thermal management system for the whole vehicle.
Source: Speaker material
This milestone was achieved in just two years, from the establishment of Zero One (March 2022) to the mass production of vehicles (April 2024). In addition, in May, we officially released our vehicle and intelligent driving strategy, and received honors in this field, demonstrating the company's rapid pace of development and innovation capabilities.
Mass production of autonomous driving requires software, vehicles and the right soil
Software
I would like to share our thoughts in depth in the field of autonomous driving, especially the current and future of autonomous driving and autonomous driving. This year, assisted driving has achieved large-scale mass production, and the installation rate of passenger cars is nearly 47%, indicating that the technology has penetrated into the market. However, autonomous driving is still in its infancy, and only a few companies such as Waymo and Carrot Express operate on a small scale in specific areas. Zero One aims to promote the mass production of autonomous driving, especially in the field of commercial vehicles, to achieve significant cost reduction and efficiency increase.
The development of autonomous driving involves three core elements: software, vehicles, and operating environment. As far as software is concerned, although autonomous driving technology is highly dependent on software, its business model has not yet fully followed the low marginal cost principle of the traditional software industry. At present, unmanned driving systems face high development costs and limited scenario migration capabilities, which makes it difficult to quickly replicate and expand.
Source: Speaker material
At the root of the problem lies in the complexity of today's autonomous driving systems, which consist of numerous modules, each of which needs to be developed and optimized independently. This modular design not only improves the accuracy of the system, but also increases the difficulty and cost of cross-scenario migration. Therefore, how to reduce the marginal cost of unmanned driving systems and achieve efficient replication and wide application has become a major challenge for the industry. In the process of R&D, we should not only pay attention to the breakthrough of the technology itself, but also pay attention to the optimization of the system architecture and the collaboration between modularization, so as to reduce the development cost and improve the generalization ability.
At present, artificial intelligence systems are rapidly evolving, and ZeroOne Auto is developing an end-to-end system based on large language models. The core reason for the adoption of this system lies in two points:
First, with the development of AI, from small expert models a decade ago to large language models that can understand common sense (such as ChatGPT), AI has been able to understand complex scenarios and open-ended propositions more deeply, rather than being limited to specific tasks. This shift enables the model to leverage built-in common sense to implement more complex tasks, reduce rule dependencies between modules, and enable end-to-end processing from sensors directly to regulatory instructions, greatly simplifying the system architecture.
Second, the end-to-end system replaces the traditional hundreds of modules through a single large model, and the system migration is mainly data-driven, without secondary development, which greatly reduces costs and enhances flexibility. Our preliminary results show that the model can navigate autonomously without HD maps and preset rules based on camera input and vehicle history data, and has demonstrated strong generalization capabilities in different countries (including Singapore, where driving rules are opposite), and can adapt to new environments with only data and related input adjustments.
Source: Speaker material
The security of large language models is a common concern. In our view, security issues can be addressed in a breakdown:
- Traceability: Large models may seem like "black boxes", but our models can output both perception and understanding of the world, and these outputs or intermediate results can be well traced and analyzed to enhance the interpretability of the model.
- New technology intervention: Technically, the controllability of the model can be improved through optimization algorithms, such as multimodal alignment, generative adversarial models, and reinforcement learning. We have observed significantly better results than those without these technologies.
- Validation and simulation: Validation and simulation are key components of autonomous driving safety. We use a large amount of offline data to conduct open-loop and closed-loop simulation tests, with the goal of reaching nearly 1 billion kilometers of verification mileage to ensure the safety of the system in various scenarios. This capability is especially valuable for OEMs to speed up the product validation process.
- Redundant design: In order to ensure foolproof, redundant design is implemented, including the safety system, and the large model is checked and analyzed to further improve the reliability of the system.
vehicle
When it comes to the importance of vehicles to autonomous driving, we can look back at the development of Waymo and Tesla. In the early stage of autonomous driving development, Waymo seemed to be closer to the future of high-end unmanned due to its complex and sophisticated expert system design and closed-loop scenario. At that time, the sensors and architectures of driverless and driver-assisted systems were not yet compatible, and although Tesla had abundant data, it could not help driverless systems.
Today, however, the landscape has changed. The passenger car field is undergoing an end-to-end transformation, with the value of data highlighted and the system complexity reduced, and Tesla has a significant advantage in this context. What should not be ignored is that the quality of the vehicle is crucial for autonomous driving. Tesla's large-scale accumulation in the field of vehicles far exceeds that of Waymo. For commercial vehicles, because the failure rate of commercial vehicles is much higher than that of passenger cars, vehicles have become a key element in the implementation of autonomous driving systems.
In the previous research and development of TuSimple autonomous driving, we encountered a large number of vehicle problems, and even became a bottleneck for product landing, which prompted me to return to China to found Zero One Automobile, firmly believing that solving vehicle problems through car manufacturing is the best or even the only way to unmanned driving.
Source: Speaker material
The key to mass-produced autonomous driving is to rely on mass-produced models as a platform. Through the natural sales of production models, we have realized the shadow mode and data closed-loop, which significantly reduces the cost of data acquisition for autonomous driving. To do this, we have built two core systems:
- The "Zhizi" system built around the ultimate data application: integrates the concept of autonomous driving and vehicle research and development, monitors thousands of signals in the whole vehicle, comprehensively captures and monitors vehicle abnormalities, and realizes real-time monitoring and problem tracking. This will not only improve the performance of autonomous driving, but also significantly improve the quality of production vehicles, forming a good closed loop and complementarity.
- The underlying vehicle platform and drive-by-wire system "Water Drop": Specially designed for unmanned driving, in cooperation with industry leading enterprises, based on a new tram architecture. The system far surpasses traditional fuel vehicles in terms of responsiveness, operational accuracy, latency and reliability, and is also the most advanced set of certified architecture in China, especially in the field of electric vehicles. Thanks to our core design and forward development strategy, we ensure a comprehensive upgrade of vehicle performance.
soil
That is, the commercialization path of unmanned driving. At present, autonomous driving is moving from technology to product and commercialization, and the main challenge lies in the breadth and complexity of business scenarios, especially passenger cars and Robotaxi, which need to realize services between any two points in a large range, involving huge asset investment, multi-party coordination and public acceptance. In contrast, commercial vehicles show unique commercialization advantages due to their relatively fixed scenarios and controllable asset scale.
Source: Speaker material
Combined with Zero-One's autonomous driving platform and strong generalization algorithm, we conceived a new ecological model: through data-driven scenario generalization, zero code delivery can be achieved and secondary development costs can be reduced; At the same time, with the use of a universal vehicle platform, one system can support a variety of scenarios, disassemble large problems into multiple small problems to solve, and improve efficiency.
In summary, the future of autonomous driving lies in: having highly generalized algorithms to adapt to changing scenarios; Solid vehicle platform as support; and scenario-specific applications that can quickly close the loop. This is at the heart of our thinking about the rapid development path of autonomous driving, especially in the field of commercial vehicles.
(The above content is from the keynote speech "The Road to Mass Production of High-end Intelligent Driving" delivered by Huang Zehua, founder & CEO of Zero One Automobile, at the 4th Commercial Vehicle Autonomous Driving Conference of Gasgoo on September 13, 2024.) )