Guosheng: General-purpose models cross the threshold for embodied manipulation; technology highlights the investment value of contact-level perception and control segments.

Guosheng released a research report stating that it recommends paying attention to hand-eye-brain integration, dexterous hands, tactile sensing, and micro-motor module segments. The essence of the physical interaction brain is to fill the control vacuum between general-purpose large models and low-level physical execution, pushing Siasun Robot&Automation from "able to demonstrate" to "able to work." The emergence of hand-eye-brain integration technology once again highlights the investment value of contact-level perception and control segments in the industry chain. Recommendations: 1) hand-eye-brain integration; 2) tactile sensing; 3) dexterous hands & micro-motor module-related targets. Guosheng's main points are as follows: GPT-6 Astra marks general-purpose models crossing the threshold for embodied manipulation, but there remains a cliff in fine contact control at the "last centimeter," urgently requiring a new technological paradigm independent of large models and with tactile sensing as its native input. Astra can already directly generate action commands to control robotic arms, achieving a generational leap in spatial reasoning tasks such as grasping blocks, with a success rate far exceeding previous generations. However, when faced with tasks requiring millimeter-level contact adjustment, such as puzzle insertion, its success rate drops sharply to 10%. This indicates that existing general-purpose models have a huge gap in physical control after contact, and pure spatial reasoning cannot replace contact-level control. As the physical carrier of the "last centimeter," the core pain point of dexterous hands lies in tactile feedback and real-time force control. Therefore, the industry's future breakthrough point is no longer simply expanding VLA model parameters, but rather shifting toward building a new paradigm of a physical interaction brain that takes tactile sensing as its core input and can truly solve contact-level control problems. The dexterous hand is not merely a hardware engineering matter, but the physical interaction intelligence carrier of Siasun Robot&Automation. The multi-finger, multi-joint mechanical structure and high degrees of freedom only endow Siasun Robot&Automation with a basic physical action space, which is a necessary but not sufficient condition for accomplishing complex tasks. In real unstructured environments, when faced with object slippage, deformation, or positional offset, relying solely on preset trajectories or high-DOF mechanical actuation cannot accomplish general-purpose precision assembly and in-hand manipulation. This requires dexterous hands to break through the pure actuator positioning and possess real-time perception of contact states, environmental understanding, and low-level dynamic feedback correction capabilities. Therefore, the industry logic of dexterous hands is transitioning from a pure contest of mechanical performance to integrated intelligent control of "perception-decision-execution," and its essence has evolved into a software-hardware collaborative system deeply integrating multimodal sensing and real-time algorithms. The "physical interaction closed loop" built by hand-eye-brain coordination is the core to breaking through the bottleneck of Siasun Robot&Automation's fine manipulation, and Daimon Siasun Robot&Automation's "physical interaction brain" provides key technical validation for this path. For high-contact tasks such as precision insertion and flexible body manipulation, relying solely on single-point capabilities of vision, large models, or dexterous hand hardware has reached its limits, and the industry urgently needs deep integration of "perception-decision-execution-feedback." As a typical representative of this paradigm, Daimon Siasun Robot&Automation's Daimon-TWM (Tactile-Anchored World Model), released in August 2026, integrates physical cognition, deductive decision-making, and instantaneous manipulation into a unified framework, truly creating a "physical interaction brain" that drives Siasun Robot&Automation's dexterous manipulation. Test data show that the model achieved average success rates of 64% and 53% in undisturbed and disturbed environments, respectively, for contact-intensive tasks, verifying that predictive decision-making and high-frequency closed-loop control significantly improve manipulation fault tolerance. This marks that competition in physical AI is shifting comprehensively from single hardware parameter or model scale contests to system-level collaborative barriers centered on hand-eye-brain integration. Risk warnings: Risk of Siasun Robot&Automation application model development falling short of expectations; passage-of-time risk; sample error risk.