The annual Industrial Expo is regarded as a global industrial innovation vane as a concentrated display platform for core technologies and products in the domestic and foreign industrial fields.
Under the Industrial 4.0 strategy, the wave of digital transformation, the rapid development of multimodal large models, cloud computing, and big data, the display of new quality productivity achievements such as advanced CNC machine tools, swinging robotic arms, smart and efficient industrial solutions, and intelligent digital future factories has become the focus of the just-concluded 24th Industrial Expo.ABB's new generation robot control platform OmniCore realizes the comprehensive integration of artificial intelligence, sensors, cloud computing and edge computing systems. Compared with the previous ABB controller, OmniCore increases the robot's operating speed by 25% and reduces energy consumption by 20%.The composite robot launched by Hikvision is equipped with an intelligent camera and multi-sensor data fusion perception to achieve integrated operation capabilities of movement, grasping and handling, which can simplify complex material handling scenarios and achieve more efficient industrial automation.The embodied intelligent industrial robot "Chuang TRON" jointly created by Weiyi Zhizao and Jiebot,Integrating perception, drive, control, algorithm, cloud service and other technologies, it can better understand and adapt to complex industrial environments, easily solve the flexibility and adaptability problems of production lines, and accurately and efficiently perform various diversified tasks.In the relatively noisy environment of the Industrial Expo, Chuang TRON can still capture dynamic environmental changes in real time through the embodied vision module, perform millisecond-level real-time map reconstruction, ensure the real-time interaction between the robot and the external environment, and have high-intensity complex environment perception capabilities.Under the multiple tasks given by the on-site staff, Chuang TRON does not require traditional teaching and robot programming, and accurately captures pictures, videos, actions, etc. Combined with large industrial vertical models, it can quickly realize task understanding and splitting. And Chuang TRON can quickly realize flexible tangents through high-speed real-time mass production execution, shortening the tangent time to hours.In addition, the real-time control frequency of the TRON robot arm is 1KHZ, and the scheme based on optimization and sampling is adopted, which can complete point-to-point path planning in milliseconds, and generate the optimal trajectory of distance and speed in real time to further ensure the overall real-time and accuracy of the system. With the mass listing of TRON in more factories next year, this will help the manufacturing industry save the cost of later tangents and maintenance, and will continue to promote the transformation and upgrading of the manufacturing industry towards intelligence.Not only at the Industrial Expo, but also at the 2024 World Robot Conference held earlier, embodied intelligence and large models have also become the focus of the market. As an important carrier of embodied intelligence, the robot arm is showing a strong prospect from commercialization to the application of new technologies of embodied large modal large models.1. "Big Model + Robot" opens the "Intelligent Machinery Era", how to solve the data problem?With the rapid development of current AI large model technology, these models integrate perception, cognition and decision-making capabilities to upgrade robots from single-function execution units to intelligent systems with autonomous learning and optimization capabilities. This transformation enables robots to better adapt to complex and changing industrial environments and significantly improve production efficiency and flexibility. "Big model + robot" is opening the "smart machine era" for embodied intelligent industrial robots.Taking Figure 02, the "strongest humanoid robot on the planet", as an example, its GPT-4 multimodal big model, onboard visual language model (VLM), and big models similar to RT-X robot control, as well as hardware such as microphones and speakers, not only significantly improve its common sense reasoning ability and task execution intelligence compared to Figure 01, but also enable natural dialogue with humans. In addition, the customized AI model and VLM make Figure 02 suitable for light-load handling and sorting tasks such as industrial manufacturing and warehouse logistics, providing fully automatic intelligent decision-making and execution capabilities.However, Liu Zhiyi, a leading Chinese artificial intelligence scientist, researcher at Qingyuan Research Institute of Shanghai Jiaotong University, and member of the Embodied Intelligence Committee of the China Artificial Intelligence Society, pointed out that in the model training of embodied intelligent industrial robots, the training data comes from multiple challenges:The first is the quality and diversity of data. The complexity of industrial scenarios requires the model to be able to understand and handle various situations, requiring a large amount of high-quality and diverse data, and the data must cover normal operations, abnormal situations, different equipment types and various production processes. Obtaining this data has become a major challenge.The second is the problem of professional data labeling. Industrial data is highly professional and requires experts with a deep industrial background to label. These expert resources are scarce, the labeling process is time-consuming and costly, and ensuring the consistency and accuracy of the labeling is the key to improving model performance.The third is the real-time requirement of data. The industrial production environment changes dynamically, and the model needs to be continuously updated and iterated with the latest data. This requires the establishment of an efficient data collection, processing and model update process to achieve near-real-time data collection and model update without affecting production.The fourth is the data security issue. Industrial data often involves core corporate secrets, such as production processes, equipment parameters, etc. How to achieve data sharing and model training under the premise of protecting privacy and intellectual property rights is a complex problem that requires both technical and management innovation.Data issues also bring technical challenges. Dr. Feng Lei, dean of the Flint Creative Industry Research Institute, believes that on the one hand, industrial large models need to process multimodal data, but multimodal data modeling and explainable machine learning models are one of the current challenges. This not only involves technical difficulties, but also includes how to make the model more transparent and easy to understand.On the other hand, with the increasing complexity of industrial large model applications, a single model can no longer meet all needs, so integrated learning and multi-model collaboration have become a development trend, but this increases the complexity and difficulty of interpreting the model.
In addition, compared with general large models such as text-to-text, text-to-picture, and picture-to-picture, the factory end-to-end closed-loop ecosystem also brings new challenges to model training.
Image source: DoNews
Dr. Feng Lei pointed out that there are often data silos between different departments and systems within the factory, resulting in the inability to effectively share and utilize data. This requires factories to strengthen data governance and management, break data silos, and achieve data interconnection and sharing.Liu Zhiyi also pointed out that the factory side needs to open up the data silos of the entire process from raw materials to finished products, including supply chain, production, quality inspection, logistics and other links. The data formats and protocols between different systems and equipment may be inconsistent, and data cleaning and integration are required. Ensuring the time series integrity of data is crucial to supporting modeling and optimization of the entire production process.Taking the quality inspection link mentioned by Liu Zhiyi as an example, Weiyi CEO Zhang Zhiqi pointed out that how to solve the contradiction between the small amount of sample data and the accurate judgment of model capabilities has become a difficult problem.On the one hand, the improvement of the yield rate in the manufacturing industry has led to a small amount of sample data, which in turn affects the model training time. On the other hand, "overkill" and "missed inspection" are two important indicators for measuring the accuracy of the factory site. "Overkill and missed inspection If any indicator is too high, the customer believes that the equipment cannot be used, and internal personnel need to be arranged for a second re-inspection. What is the significance of customers using AI quality inspection equipment? If the overkill rate is controlled to less than 5%, higher requirements will be placed on sample data."Facing the training data problem, Weiyi built a closed loop of model development and data collection on the production line, deployed equipment with pre-trained models directly on the production line, and turned production line operators into "model trainers". The results generated by the model were re-judged and corrected on the "human-computer interactive model training platform" developed in the cloud, and the model then sent the corrected results to the equipment for execution.At the same time, the correction process is used as the "new sample" for the next iteration of the model, so that the model can continuously learn from the work experience of human masters, so that a "end-to-cloud" model training closed loop is formed on the production line, which enables the equipment to be "out of the box" and obtains a large amount of real-time production line data for model training.Based on its competitiveness in the AI quality inspection market and years of data accumulation, Weiyi has the world's largest unstructured industrial precision standard database. Cao Wei, managing director and partner of LanChi Ventures, pointed out in an interview with the media that excellent companies in the industrial robot industry will form their own data closed loop, hardware closed loop, and algorithm closed loop on the algorithm side. The scarcity of data and the iteration speed between data and algorithms determine the core capabilities of enterprises in this field.
Source: IDC "China AI-enabled Industrial Quality Inspection Solution Market Share 2023"
Zhang Zhiqi also pointed out that Weiyi has currently accumulated a certain amount of business and various types of data and model capabilities. Even if peers work hard for three to five years, it is difficult to do it, and even if peers do it, it may be difficult to adapt to market changes. This is one of Weiyi's core competitiveness.2. How to equip smart robots with "brains" to make them closer to humans?In addition to the problem of training data, from the perspective of the industrial robot exhibition area of this Industrial Expo, most robots are equipped with a "brain" responsible for command and debugging. Although they run according to the established program, the staff needs to check whether the operation is normal from time to time.However, from the perspective of KUKA and Fanuc used by Tesla's super factory, ABB and KUKA widely used in BMW's Dingolfing factory in Germany, and tens of thousands of Kiva used in Amazon's global warehouses, these industrial robots have sufficient advantages in driving and execution, which is also their consistent technical core.However, with the perception and cognitive capabilities brought by AI, foreign companies have not yet made great efforts to deploy, and the application of embodied intelligence in the industrial field is in a state of absence. This has given. This has given the domestic embodied intelligent industrial machine industry an opportunity. Although multiple problems need to be overcome in the short term, it has become a breakthrough point that cannot be ignored by domestic companies to give full play to the advantages of embodied intelligence in industry.Dr. Feng Lei pointed out that in terms of perception technology, it is necessary to enhance its sensor technology, including vision, touch, force and other sensors, in order to achieve accurate perception of complex industrial environments. At the same time, it is also necessary to continuously optimize algorithms to improve sensor accuracy and response speed.In terms of cognitive technology, the cognitive ability of industrial robots depends on the development of artificial intelligence technology, especially deep learning, natural language processing and other technologies. What needs to be broken through in the short term is how to effectively apply these technologies to industrial robots so that they can understand and handle complex industrial tasks and make autonomous decisions.In terms of drive technology, drive technology includes key components such as servo motors and reducers, which are the core of industrial robots. Chinese industrial robots still have certain technical bottlenecks in these key components, and need to strengthen independent research and development to improve performance and stability.In terms of execution technology, execution technology is directly related to the operating accuracy and efficiency of industrial robots. It is necessary to continuously optimize control algorithms, improve the robot's motion control ability and accuracy, and ensure safety and reliability during execution.Liu Zhiyi also pointed out that at present, core components such as high-precision sensors, controllers, and servo motors of industrial robots still rely heavily on imports. In addition, the deep integration of software and hardware is a systematic challenge. Enterprises need to break through the traditional mechanical design thinking and consider the coordination of software and hardware from the system level. Develop more flexible modular designs, support rapid function customization and upgrades, realize software-defined hardware, and improve the adaptability and reconfigurability of robots.General AI algorithms need to be optimized for industrial scenarios to improve accuracy, stability and real-time performance. Develop more efficient edge computing algorithms to achieve localized intelligent decision-making. At the same time, improve the interpretability and reliability of algorithms to meet the strict requirements of industrial applications.Zhang Zhiqi also pointed out that Weiyi's technical strategy of "eye-hand-brain cloud" has faced the problem that the visual system is like the eyes, and the motion mechanism of the robotic arm is similar to human hands and feet. The gap between the two is like a blind man carrying a lame man, and the lame man is directing the blind man to move forward and backward. There are many difficulties in actual implementation, and it is often necessary to continuously stack manpower to make up for the problems between them. This should be a complete system. The integration of perception, cognition, planning, driving, and control capabilities is the fundamental way to solve such problems.Based on this, Weiyi established a special embodied intelligence project team in 2023, and jointly developed with domestic robot manufacturer Jiebot to achieve a breakthrough in the integration of industrial AI and industrial robots, and break through the barriers of software and hardware quality inspection.If the problem of "data" and "smarter" is a technical problem, the problem above technology is how to make embodied intelligent robots better meet the needs of downstream customers.
Source: DoNews
In recent years, with the gradual diversification and personalization of consumer demand, the challenges of supply chain and market uncertainty that Chinese companies need to face when going overseas, the integration of globalization and customized demand, and the manufacturing industry's increasing pursuit of improved resource utilization efficiency, reduced production costs and improved innovation capabilities, the global manufacturing industry is increasingly pursuing flexible production on the factory side, and it has become an irreversible trend. The overseas industrial robots mentioned above can "enter the factory to work" based on meeting the needs of these companies for flexible production.However, compared with overseas markets, the ability of domestic industrial robots to cope with flexible production in the manufacturing industry still needs to be improved. Zhang Zhiqi pointed out that traditional industrial robots move in a confirmed and closed space and can only perform single repetitive actions. They have weak generalization ability and cannot have flexible tangent ability, which limits the popularization and application of industrial robots from the source.Faced with the actual demand of the manufacturing industry for extreme cost and flexible production, embodied intelligent robotic arms launched by manufacturers including Weiyi, Aobo Robot, Tosda, Siasun Robot, Efort Intelligent Equipment, Huazhong CNC, ABB China, etc. are equipped with programming-free functions.The advantage of programming-free robotic arms is that they can quickly adapt to different production tasks through adaptive technology, especially in small batch and multi-variety production environments. They can automatically learn new operation paths and quickly put them into use according to different process and task requirements. The robotic arms perceive the environment through visual sensors, force sensors, etc., and automatically adjust the operation actions based on machine learning technology, which can achieve a higher level of flexible production without reprogramming every time the task changes, thereby improving factory production efficiency.Cao Wei pointed out that the next development trend of industrial robots is to move towards light delivery and intelligence, to make delivery light, and most importantly, no programming. Although industrial robot programming may still need customization, it can understand what I want to customize during the conversation with the robot, rather than continuing to ask people to write code.Zhang Zhiqi pointed out that the embodied intelligent robots created by Weiyi can not only occupy the existing market, but also enter flexible and flexible scenarios that traditional industrial robots cannot enter, helping industrial manufacturing companies solve the problems of slow tangent and low efficiency when using traditional industrial robots.3. The speed of embodied intelligent industrial robotic arms landing may be faster than that of humanoid robots
In addition to meeting the flexible production of the manufacturing industry, the cost account and economic account of labor costs of industrial robots are actually the most concerned things for business owners. The prices of overseas industrial robots vary greatly due to different industries, industry needs, functions and load capacities.
Generally speaking, the price of basic industrial robots is about 20,000 to 80,000 US dollars, and high-end intelligent robots may reach 100,000 to 300,000 US dollars or even higher. However, the high profits and high labor costs of high-end manufacturing industries in Europe and the United States make the "comprehensive cost performance" advantage of intelligent robots outstanding. In contrast, domestic industrial robots have an absolute advantage in price. This advantage is not only the key to domestic and foreign competition, but also the key for domestic companies to bring "intelligence" into factories.
Source: DoNews based on public information
Humanoid robots, which are very aggressive in momentum, have high costs because they use many expensive components. Just talking about visual perception, many expensive components are used, such as 3D cameras, laser radars, etc.
For the domestic manufacturing industry, their profit level is relatively low and the price war that has affected many industries in the past two years has continued to impact the profits of the manufacturing industry. It is not realistic to expect business owners to pay for humanoid robots that cost millions of dollars, not to mention that they will pay attention to what kind of large model technology and capabilities the humanoid robots use.
Moreover, the cost of using industrial robots in the manufacturing industry not only includes the initial purchase cost of the enterprise, but also includes the subsequent maintenance and upgrade costs and the cost of connecting the data of the enterprise's internal MES, ERP, and supply chain management systems, the overall solution cost, and the coordination cost of industrial robots and humans working together.
If industrial robots really enter the factory to "work", the factory not only needs to add additional safety fences and sensors to ensure the safety of workers. And if the robot fails, in addition to increasing the new maintenance cost, it will also affect the production of the production line, which will affect the factory's delivery time to downstream customers, and then increase costs. Based on this, public data shows that the penetration rate of robots per 10,000 industrial workers in China is only more than 392, which is at a relatively low level.
Liu Zhiyi pointed out that in order to increase the penetration rate of domestic industrial robots, systematic measures such as optimizing product design, developing service-oriented business models to lower the threshold for use, using AI to improve efficiency is an innovative direction, optimizing robot structure and control system through AI-assisted design, and encouraging enterprises to adopt domestic industrial robots through tax incentives and subsidies to reduce costs, but promoting the localization of core components is the fundamental way to reduce costs. Through industry-university-research cooperation, breakthroughs in key technologies such as high-end reducers and servo systems. Establish a localized supply chain system, improve the standardization and versatility of components, and achieve large-scale production.
Dr. Feng Lei also pointed out that technological innovation and independent research and development are the key to reducing costs. On the one hand, strengthen the research and development of core technologies and promote the localization of core components such as servo motors and reducers. Through independent innovation and technological breakthroughs, reduce dependence on imported components, thereby reducing costs from the source.
On the other hand, through technological innovation to improve the performance of industrial robots, such as improving accuracy, enhancing stability, optimizing algorithms, etc., make products more competitive, thereby gaining higher recognition and market share in the market.
It should be pointed out that compared with humanoid robots, embodied intelligent robotic arms not only have cost advantages, but also have clearer and more extensive application scenarios in industrial automation, logistics, service industries and other fields, which means that the large-scale landing speed of embodied intelligent robotic arms will be higher than that of humanoid robots in the future.
Cao Wei pointed out that the outstanding problem of industrial robots at this stage is that although everyone has business to do, they are not making money due to high delivery costs, too many personalized things and low versatility. Based on this, the most worthy issue of industrial robots is how to improve delivery efficiency, reduce delivery costs and make it ready to use out of the box, so as to achieve the real intelligence of industrial robots,
If industrial robots can achieve more intelligent out-of-the-box use, or more intelligent lightweight delivery. From project-based to product-based, industrial robots will be a profitable and fast-growing supermarket in the future.
It is worth noting that Weiyi's "Chuang TRON" can be used out of the box and delivered in a lightweight manner. This is why Zhang Zhiqi mentioned that it is expected that within 1-2 years, "Chuang TRON" will further promote the expansion of the market scale of embodied intelligent industrial robots. Embodied intelligent technology will make the deployment of industrial robots more agile. In the future, the market scale of industrial robots will further expand to 1 million to 1.5 million sets/year, and the annual output value will reach 100 billion to 150 billion.
With the continued empowerment of more industrial robots and manufacturing industries by companies including ABB, Weiyi, and Hikvision in the future, as well as the continued efforts of TOG policies, and the continuous filling of technical shortcomings by universities and enterprises, this will not only help more industrial robots "enter the factory to work", but also usher in the "iPhone moment" for the industry.
More importantly, this can help the manufacturing industry reduce costs and improve efficiency, and make the domestic manufacturing industry continue to grow and become stronger.
