Caitong: The era of physics AI is coming, with data becoming an important drive.

Caitong releases a research report stating that data is the key factor determining the development of physical AI at present. With the acceleration of the physical AI industry, the demand for corresponding data is expected to continuously increase, and orders from various data vendors are also likely to grow rapidly. The main data vendors of physical AI mainly include vendors providing data sets or simulation engines, as well as vendors producing data acquisition equipment and other hardware. 1) Key components of data acquisition equipment include cameras, IMUs, data gloves, etc. 2) Providing simulation or data asset-related targets. The main viewpoints of Caitong are as follows: Data acquisition faces some challenges, and data is the key factor determining the development of physical AI at present. The development of physical AI cannot be separated from the three elements of algorithms, computing power, and data. Currently, algorithm models are gradually converging, and data has become an important factor in determining the development of physical AI. Unlike digital AI, physical AI has higher data requirements, such as multimodal coverage, spatial and temporal alignment, and high physical accuracy. At present, data acquisition in physical AI faces challenges such as "small scale, high cost, lack of quality standards, and a large gap between virtual and reality." The competition among physical AI vendors has shifted from competing in algorithms and computing power to seeking the ability to supply high-quality, large-scale data. Different types of data for physical AI have obvious advantages and disadvantages, and data fusion has become a trend. The types of data for physical AI form a "pyramid" structure, with massive internet public information at the bottom layer; simulation data and data obtained from data acquisition without ontology (UMI, Ego, etc.) in the intermediate layer; and real machine data such as remote control at the top layer. Various types of data have obvious advantages and disadvantages, and a single type of data cannot simultaneously meet the requirements for "authenticity/precision, scale, cost, efficiency," etc. The industry is gradually merging various types of data, leveraging the advantages and disadvantages of each, such as using simulation data to increase data scale and cover long-tail scenarios, and fine-tuning high-quality data such as real machine data. The government attaches importance to data development, and various data vendors have a rich variety of products. The Chinese government attaches importance to the development of data in the field of physical AI and has issued relevant documents to support it; Siasun Robot&Automation, a human-shaped intelligence training center in Beijing, Shanghai, and other places, has also established data training grounds. At the vendor level, many data vendors have launched different products. For example, vendors such as Deerbright Siasun Robot&Automation and MiFeng Technology have released data acquisition devices without ontology (UMI/Ego), while vendors like Guangluo Intelligence mainly provide simulation synthetic data as the main source and offer data engines. In addition, some vendors are introducing open-source data sets or launching data malls to collectively promote industry acceleration. Risk Warning: Slow progress in the physical AI industry, intensified vendor competition, and tight resources of infrastructure (storage, etc.).