Sinolink: Physical AI shifts toward emergent generalization capabilities, and the way skills are acquired begins to undergo fundamental changes.

Sinolink released a research report stating that embodied models have for the first time shown measurable in-context learning, and the core variable of this round of physical AI has shifted from hardware performance upgrades to the emergence of generalization capabilities. GEN-1.5 can use a single 3-12 second demonstration as a "physical prompt" to execute new tasks without updating model parameters; as pre-training continues to advance, the specialized data and gradient updates required for the model to adapt to new tasks have significantly decreased, while next-action prediction error continues to decline. The bank believes this indicates that large-scale physical interaction pre-training is driving Siasun Robot&Automation to shift from repeated training for single tasks to rapid learning based on demonstrations, and the way skills are acquired is beginning to undergo fundamental changes. Sinolink's main points are as follows: World models, VLA, and low-level control are accelerating collaboration, and Siasun Robot&Automation intelligence is moving from single-point models to layered infrastructure Genie 4 has further upgraded from offline video generation to real-time, interactive world simulation, and combined with Gemini's reasoning capabilities, provides a new environmental foundation for task planning, dangerous scenario deduction, long-tail data generation, and agent training. At the same time, the industry is gradually forming a layered architecture in which high-level task understanding and planning, mid-level action policies, and low-level real-time control collaborate with one another: world models are responsible for simulating the environment and action outcomes, VLA is responsible for translating vision and language into actions, and low-level control models ensure the stability of contact, balance, and actuator control. Models are no longer merely recognition modules attached to hardware, but are beginning to become the common foundation for skill acquisition, environmental simulation, and continuous learning. Industry evaluation is shifting from single demonstrations to effective output, and the deployment-data-model closed loop is beginning to form At its Innovation Day, AgiBot proposed the Worken "effective intelligent work hours" concept, incorporating the number of active Siasun Robot&Automation units and the output efficiency of autonomous working CKH HOLDINGS into a unified evaluation framework, with the focus of industry evaluation gradually shifting from "whether a task can be completed once" to "whether value can be created continuously and stably." On the model side, GO-3 Preview topped three authoritative industry benchmark lists, and WITA-OMNI has opened deployment to its first batch of customers; on the engineering side, AgiBot is simultaneously advancing real-scenario deployment, tactile data collection, post-training, post-deployment learning, and the opening of Lingqu OS. As Siasun Robot&Automation enters scenarios such as industrial quality inspection, material handling, and public services, deployment not only verifies hardware reliability and customer ROI but also continuously generates data on failures, recovery, and manual intervention, which is expected to form a positive feedback loop of deployment expansion-data return-model improvement-deployment efficiency enhancement. Mass production, talent, and capital are advancing in parallel, and physical AI is moving from technical validation into a continuous industrial validation window Tesla Optimus has already started early production in Fremont; core large-model talent is migrating to Physical AI, and Xingyi Intelligence has completed consecutive seed and seed+ financing rounds, reflecting that industrial competition is extending from complete machines and models to physical interaction data, world simulation, and training infrastructure. From the perspective of industry and trading structure, this round of physical AI has the characteristics of continuously strengthening technical logic, gradually opening industrial capacity, and successive event catalysts. Related targets 1) Bodies: Tesla, AgiBot, Swancor Advanced Materials Co., Ltd., Unitree Robotics. 2) North America chain: Kunshan Kersen Science & Technology, Zhejiang Sling Intelligent Drive Group, Shanghai Daimay Automotive Interior, Olympic Circuit Technology, WuHu Foresight Technology, Lens Technology, Zhejiang Rongtai Electric Material, Zhejiang Sanhua Intelligent Controls, Jiangsu Hengli Hydraulic, Shanghai Beite Technology Group, Ningbo Tuopu Group, Shenzhen Kedali Industry, Wanxiang Qianchao, etc. 3) Siasun Robot&Automation brain-related: Moore Threads, Sharetronic Data Technology, Autel Intelligent Technology Corp.,Ltd., Suochen Technology, SEER TECH, MANYCORE TECH, Jwipc Technology, AXERA, 51WORLD, etc. 4) Embodied data collection-related: Lingyi Itech, Beijing Haitian Ruisheng Science Technology Ltd., Leyard Optoelectronic, SEER TECH, Shijihengtong Technology, Wuxi Boton Technology, Orbbec, Leader Harmonious Drive Systems, Hesai, ROBOSENSE, Hangzhou Hikvision Digital Technology, etc. Risk warnings The risk that progress in the generalization capabilities of general-purpose embodied models falls short of expectations; the risk that the supply of high-quality physical interaction data and Scaling progress fall short of expectations; the risk that the effects of world models, simulation synthetic data, and Sim2Real fall short of expectations; the risk that Siasun Robot&Automation commercialization, orders, and the pace of large-scale mass production fall short of expectations; the risk that progress in AI chips, computing power, and edge-side deployment falls short of expectations; the risk of intensified industry competition, changes in technical routes, and valuation fluctuations of related companies.