TU-TESL

• 22

Alibaba's layout started at the end of 2015 and lasted for four years. Its terminal distribution robot has landed on a large scale and entered the university campus to undertake the task of package delivery. Last year's double 11, these unmanned delivery vehicles also joined in the battle, receiving and sending more than 1000 packages a day on a single campus. It is not hard to find that, unlike most enterprises, Alibaba's layout of automatic driving is a breakthrough from terminal logistics. Wang Gang, head of Alibaba Da Mo yuan's automatic driving laboratory, told auto heart that Alibaba's goal of automatic driving is to build an intelligent logistics transportation platform to make logistics more convenient and efficient. The establishment of this goal is closely related to the attributes of Ali economy. Taobao, tmall, hungry, HEMA and other businesses are inseparable from logistics and distribution. Ali also has a smart logistics platform such as rookie network, which has a deep understanding of the industry and many cooperative logistics enterprises. Therefore, it is natural for Ali to start from the logistics industry in order to realize the commercial landing of automatic driving technology. The safety risk of using the self driving vehicle to carry things is lower than that of carrying people, the difficulty of technology realization is also lower, and there is more freedom in laws and regulations. After positioning to build an intelligent logistics transportation platform, Ali has also established a strategy of two business forms of end-to-end unmanned distribution and open road logistics. Among them, the R & D Progress of end-to-end unattended distribution is faster, because of its slow vehicle speed, low hardware and software accuracy and stability requirements, and it can also be remotely controlled. At present, Alibaba's unmanned distribution vehicles have started small-scale operation in many campuses and parks. In order to achieve the goal of "building an intelligent logistics transportation platform" and promote the commercial implementation of automatic driving technology, Alibaba must have a complete development map of automatic driving technology. First of all, we focus on the algorithm layer of the big picture of Ali's automatic driving technology. Wang Gang has a point of view: at present, the biggest bottleneck restricting the development of automatic driving is still that the algorithm is not good enough, so even if the most advanced sensors and computing units in the world are integrated into a car, the car still cannot achieve full automatic driving. As a result, Ali has put more effort into the research and development of automatic driving algorithm and proposed the concept of "small front desk, large middle platform". "Small front desk" refers to the automatic driving algorithm modules such as perception, positioning, decision-making and control, which must be developed by all automatic driving R & D enterprises; and "big middle platform" refers to the autodrive independently built by Alibaba team Platform, which is composed of automatic parameter adjustment module, network structure search module, active learning module, framework and basic cluster platform, can greatly improve the speed of research and development iteration of automatic driving technology. If the autopilot task of the vehicle is compared to a tough battle, the "small front desk" plays the role of the Stormtrooper, while the "big center" is the subsequent aircraft and tank formation. "Big, middle and Taiwan" will provide strong support for "small front desk". At present, there are still a lot of artificial design links in the whole research and development link of automatic driving algorithm, such as data preprocessing, neural network structure / super parameters of perception module, fusion parameters in positioning module, rules and parameters in decision module, etc. These artificial design links limit the progress of algorithm research and development to a large extent, which makes algorithm researchers spend a lot of time to adjust parameters, poor quality and low efficiency. In order to reduce manual design, Alibaba's autodrive platform can automatically learn better network structure / parameters / data preprocessing and so on by searching / optimizing based on massive automatic driving data, so as to realize calculation instead of manual. Different from the industry's automl, autodrive can self-study based on the complex multi-modal sequential data, and can serve the algorithm module of the whole automatic driving link, including perception, decision-making planning and positioning. For example, in the face of some typical identification and detection tasks, if a detection network is designed manually, it may bring redundancy because it does not know which part is the most core network, but after the optimization of autodrive platform, the network complexity will be greatly reduced. Because of the high real-time requirements of automatic driving, reducing the network complexity can improve the overall efficiency and reduce the dependence on hardware. Behind autodrive, Alibaba has built its own cloud platform for automatic driving, and massive data (scene database, automatic driving vehicle data, data collection vehicle data) have been moved to Alibaba cloud. This cloud platform includes data management platform, automatic driving simulation platform and algorithm model training platform. Relying on these platforms, Alibaba's automatic driving team has opened a whole set of systems including data collection, data annotation, simulation, model training and evaluation, so as to make the research and development of automatic driving algorithm more efficient. At this stage, the data used by autodrive platform mainly comes from Alibaba's automatic driving test operation scene vehicles and special data collection vehicles, as well as data generated through simulation system editing. At present, the autodrive platform has been used in Alibaba's autopilot team, and its autopilot decision-making planning team, perception team and positioning team have begun to use the platform. Alibaba believes that in the future, the middle platform similar to autodrive will become an essential module for in-depth research and development of automatic driving. In addition to paying attention to algorithm research and development, Ali also has a layout at the level of autopilot hardware. Ali has invested in Suteng juchuang in the field of lidar through rookie network, and the two sides have cooperated to do a lot of customized development. In the field of cameras, Ali has customized the ISP for low illumination scenes such as night, forming a complete ISP IP. Compared with ISP, which is a general camera in the industry, it greatly improves the image quality and the perception ability of automatic driving in low illumination scene. Ali is also developing the software side of the embedded computing platform, including hardware and software co design based on FPGA and embedded software design. In fact, Alibaba has made a wide layout in the field of AI chips. Alibaba invested in Shenjian technology through ant financial before being acquired by Xilinx, which is the representative enterprise of FPGA. In the future, Alibaba's chip R & D capability should also be able to provide assistance to its autonomous driving R & D. At the sensing end, Ali's autopilot is a multi-sensor fusion scheme, including lidar, camera, millimeter wave radar, inertial navigation, etc. However, it also has its own unique features: Alibaba autopilot multi-sensor fusion system adopts the design idea of on-demand perception enhancement, which can feed back the online adaptive switching model and information fusion strategy according to the external environment and downstream decision-making planning under the limited computing resources, which can well relieve the pressure of the computing unit. In order to ensure the safety and stability of vehicles, there are many redundant designs in the system architecture of Ali automatic driving. In addition to the traditional autopilot brain, Ali designed a safety cerebellum system for his vehicle, which focuses on passive safety. In addition, Ali has introduced a set of remote driving system, which can control vehicles in dangerous situations through 5g and other communication technologies. The redundancy design is also reflected in the platform of unmanned distribution vehicles designed by Ali. These vehicles adopt a highly integrated EE architecture, which is divided into chassis domain and automatic driving domain. Each domain also has multi-layer redundancy guarantee. Alibaba will now focus on improving its single car intelligence, including in-depth research on autonomous driving algorithms, sensors and computing platforms. Following a clear commercialization goal, Alibaba's research and development of autonomous driving has become more focused. For Alibaba, its immediate goal is to build a safe, intelligent and low-cost self driving vehicle, which is also the only way to the best combination of self driving technology and business.