Alibaba's Journey to Autonomous Driving

Alibaba's layout of autonomous driving began at the end of 2015. After four years, its terminal delivery robots have been deployed on a large scale and entered university campuses to take on the task of package delivery.

Last year on Double Eleven, these unmanned delivery vehicles also joined the battle, with more than a thousand parcels being received and delivered on a single campus in a single day.

It is not difficult to find that, unlike most companies, Alibaba’s layout of autonomous driving starts from terminal logistics.

Wang Gang, head of Alibaba DAMO Academy's autonomous driving laboratory, told Auto Heart that Alibaba's goal of autonomous driving is to build a smart logistics and transportation platform to make logistics more convenient and efficient.

The establishment of this goal is closely related to the attributes of the Alibaba economy. Businesses such as Taobao, Tmall, Ele.me, and Hema are all inseparable from logistics and distribution.

Alibaba also has smart logistics platforms such as Cainiao Network, which has a deep understanding of the industry and has many cooperating logistics companies. Therefore, it is natural for Alibaba to start from the logistics industry to commercialize autonomous driving technology.

Using self-driving vehicles to carry goods has lower safety risks than carrying people, is less difficult to implement technically, and has greater freedom in terms of laws and regulations.

After positioning itself as a smart logistics and transportation platform, Alibaba has also established a strategy of advancing two business lines: terminal unmanned delivery and open road logistics.

Among them, the research and development of terminal unmanned delivery is progressing faster, because its vehicle speed is slow, the requirements for software and hardware precision and stability are low, and it can also be remotely controlled. At present, Alibaba's unmanned delivery vehicles have started small-scale operations in multiple campuses and parks.

If Alibaba wants to achieve its goal of "building a smart logistics and transportation platform" and promote the commercialization of autonomous driving technology, it must have a complete autonomous driving technology development blueprint.

We first focus on the algorithm layer of Alibaba's autonomous driving technology.

Wang Gang has a point of view: the biggest bottleneck restricting the development of autonomous driving is still the poor quality of the algorithm. Therefore, even if the most advanced sensors and computing units in the world today are integrated into a car, this car still cannot achieve fully autonomous driving.

For this reason, Alibaba has invested more energy in the research and development of autonomous driving algorithms and proposed the concept of "small front-end, large middle-end".

The "small front desk" refers to autonomous driving algorithm modules such as perception, positioning, decision-making, and control. These are algorithms that all autonomous driving R&D companies must develop; the "big middle desk" is the AutoDrive platform independently built by the Alibaba team. This platform consists of an automatic parameter adjustment module, a network structure search module, an active learning module, a framework, and a basic cluster platform, which can greatly improve the speed of autonomous driving technology R&D iteration.

If the vehicle's autonomous driving task is compared to a tough battle, then the "small front desk" plays the role of the assault team, while the "big middle desk" is the subsequent aircraft and tank formation. The "big middle desk" will provide strong support for the "small front desk".

At present, there are still a lot of manually designed links in the entire autonomous driving algorithm development chain, such as data preprocessing, neural network structure/hyperparameters of the perception module, fusion parameters in the positioning module, rules and parameters in the decision module, etc. These manually designed links have greatly restricted the progress of algorithm development, requiring algorithm developers to spend a lot of time adjusting parameters, resulting in poor quality and low efficiency.

In order to reduce manual design, Alibaba's AutoDrive platform can automatically learn better network structures/parameters/data preprocessing, etc. based on massive autonomous driving data through search/optimization, thereby replacing manual work with computing.

What makes AutoDrive different from the industry's AutoML is that it performs self-learning based on complex multimodal time-series autonomous driving data, and can serve the algorithm modules of the entire link of autonomous driving, including perception, decision planning and positioning.

For example, when facing some typical recognition and detection tasks, if a detection network is designed manually, redundancy may occur because we do not know which parts are the core of the network. However, after optimization by the AutoDrive platform, the network complexity will be greatly reduced. Because autonomous driving has very high real-time requirements, reducing network complexity can improve overall efficiency and reduce dependence on hardware.

Behind AutoDrive, Alibaba has also built its own autonomous driving cloud platform, and massive amounts of data (scene database, autonomous driving vehicle data, data collection vehicle data) have been moved to Alibaba Cloud.

This cloud platform includes a data management platform, an autonomous driving simulation platform, and an algorithm model training platform. Relying on these platforms, Alibaba's autonomous driving team has opened up a complete set of systems including data collection, data labeling, simulation, model training, and evaluation, making the research and development of autonomous driving algorithms more efficient.

At present, the data used by the AutoDrive platform mainly comes from Alibaba's autonomous driving trial operation scenario vehicles and special data collection vehicles, as well as data generated by editing the simulation system, etc.

Currently, the AutoDrive platform has been used within Alibaba's autonomous driving team, and its autonomous driving decision-making planning team, perception team, and positioning team have begun to use this platform. Alibaba believes that in the future, a middle platform similar to AutoDrive will become an essential module for in-depth research and development of autonomous driving.

In addition to focusing on algorithm research and development, Alibaba also has a layout in the autonomous driving hardware level.

Alibaba invested in RoboSense in the LiDAR field through Cainiao Network, and the two parties cooperated to carry out a lot of customized development.

In the field of cameras, Alibaba has customized the design of ISP for low-light scenes such as nighttime, forming a complete ISP IP. Compared with the current common automotive-grade camera ISP in the industry, it greatly improves the image quality and autonomous driving perception capabilities in low-light scenes.

Alibaba is also conducting research and development on the software side of embedded computing platforms, including FPGA-based hardware and software co-design and embedded software design.

In fact, Alibaba has already made extensive arrangements in the field of AI chips. Alibaba previously invested in DeePhi Technology through Ant Financial, which was later acquired by Xilinx, a representative company of FPGA. In the future, Alibaba's chip research and development capabilities should also be able to provide support for its autonomous driving research and development.

On the perception side, Alibaba's autonomous driving adopts a multi-sensor fusion solution, including lidar, cameras, millimeter-wave radar, inertial navigation, etc.

But it also has its own unique features: Alibaba's autonomous driving multi-sensor fusion system adopts an on-demand perception enhancement design concept, which can adaptively switch models and information fusion strategies online based on the external environment and downstream decision-making planning feedback under limited computing resources, which can effectively alleviate the pressure on the computing unit.

In order to ensure the safety and stability of the vehicle, Alibaba's autonomous driving system architecture also has many redundant designs.

In addition to the traditional autonomous driving brain, Alibaba has designed a safety cerebellum system for its vehicles, which focuses on passive safety. In addition, Alibaba has also introduced a remote driving system that can control vehicles in dangerous situations through communication technologies such as 5G.

Redundant design is also reflected in the unmanned delivery vehicle platform designed by Alibaba. These vehicles adopt a highly integrated EE architecture, which is divided into chassis domain and autonomous driving domain, and each domain also has multiple layers of redundancy protection.

Alibaba is now focusing its efforts on improving the intelligence of its vehicles, including deepening its research on autonomous driving algorithms, sensors, and computing platforms. Following a clearer commercialization goal, Alibaba's autonomous driving research and development has become more focused.

For Alibaba, its immediate goal is to build a safe, intelligent and low-cost autonomous driving vehicle, which is also the only way to achieve the best combination of autonomous driving technology and business.