Guojin Securities: Physical AI is shifting to generalized capabilities, and skills acquisition methods are beginning to fundamentally change
The Zhitong Finance App learned that Guojin Securities released a research report saying that for the first time, the embodied model showed measurable contextual learning, and the core variables of this round of physical AI were shifting from hardware performance upgrades to generalization capabilities. GEN-1.5 can use a single demonstration of 3-12 seconds as a “physical reminder” to perform new tasks without updating model parameters; as pre-training continues to advance, the dedicated data and gradient updates required for the model to adapt to the new task are significantly reduced, and the next action prediction error continues to decline. According to the bank, this shows that large-scale physical interaction pre-training is driving robots from repeated training for a single task to rapid learning based on examples, and skills acquisition methods are beginning to fundamentally change.
Guojin Securities's main views are as follows:
World models, VLA and underlying control accelerate collaboration, and robot intelligence is moving from a single-point model to hierarchical infrastructure
Genie 4 was further upgraded from offline video generation to real-time, interactive world simulation, and combined with Gemini's reasoning capabilities to provide a new environmental foundation for mission planning, dangerous scenario deduction, long-tail data generation, and agent training. At the same time, the industry is gradually forming a hierarchical architecture where high-level task understanding and planning, mid-level action strategies, and lower-level real-time control cooperate with each other: the world model is responsible for deducing environmental and action results, VLA is responsible for transforming vision and language into action, and the underlying control model ensures the stability of contact, balance, and actuator control. Models are no longer just recognition modules attached to hardware, but are beginning to become a common foundation for skill acquisition, environmental deduction, and continuous learning.
Industry evaluation moved from a single demonstration to effective output, and a closed loop of deployment, data, and model began to form
Zhiyuan proposed Worken's “effective intelligent working hours” concept on Innovation Day, incorporating the number of active robots, autonomous working time, and output efficiency into a unified evaluation framework. The focus of industry evaluation gradually changed from “whether it can complete a task once” to “whether it can create value sustainably and stably.” On the model side, GO-3 Preview topped the three authoritative industry evaluation lists. WITA-OMNI has opened deployment to the first batch of customers; on the engineering side, Zhiyuan simultaneously promotes real scene deployment, tactile data collection, post-deployment learning, and the opening of Lingqu OS. As robots enter scenarios such as industrial quality inspection, material handling, and public services, deployment not only verifies hardware reliability and customer return on investment, but also continues to generate failure, recovery, and manual takeover data, which is expected to form positive feedback on deployment expansion - data return - model improvement - deployment efficiency improvement.
Mass production, talent and capital advance simultaneously, and physical AI moves from technical verification to a continuous industrial verification window
Tesla Optimus has begun early production in Fremont; the core talents of the big model have migrated to Physical AI, and Xingyi Intelligence has completed the seed round and the continuous seed+ round financing round, which reflects the extension of industry competition from complete machines and models to physical interactive data, world simulation, and training infrastructure. Looking at the industry and transaction structure, this round of physical AI has the characteristics of continuous strengthening of technical logic, gradual opening up of industrial capacity, and catalytic continuation of events.
Related targets
1) Main body: Tesla, Zhiyuan, Shangwei New Materials, Yushu Technology. 2) North American chain: Keson Technology, Siling Smart Drive, Daimei Co., Ltd., Shiyun Circuit, Forsai Technology, Lansi Technology, Zhejiang Rongtai, Sanhua Intelligent Control, Hengli Hydraulic, Beite Technology, Tuopu Group, Kodali, Wanxiang Qianchao, etc. 3) Robotic brain related: Moore threads, co-creation data, Daotong Technology, Suochen Technology, Xiangong Intelligence, Cluster Nuclear Technology, Zhiwei Intelligence, Aixin Yuanzhi, May Day Vision, etc. 4) Related data collection: Lingyi Intelligent Manufacturing, Haitian Ruisheng, Riyadh, Xiangong Intelligence, Century Hengtong, Baotong Technology, Obi Zhongguang, Green Harmonics, Hesai, Suteng Juchuang, Hikvision, etc.
Risk Alerts
The risk that generalization capabilities of generic embodied models will fall short of expectations; the risk that the supply of high-quality physical interaction data and scaling will fall short of expectations; the risk that the world model, simulation synthesis data, and Sim2Real results fall short of expectations; the risk that the pace of commercialization, orders, and large-scale mass production of robots falls short of expectations; the risk that AI chips, computing power, and end-side deployment will not progress as expected; the risk of increased industry competition, changes in technology routes, and fluctuations in valuations of related companies.