As price competition intensifies across the humanoid robot hardware sector and the homogeneity of physical bodies becomes more pronounced, the industry's competitive focus is shifting decisively toward robotic intelligence systems—the "robot brain." Physical AI foundation models, serving as the core of embodied intelligence, handle environmental perception, physical world reasoning, and action decision generation, making them the decisive variable in determining whether robots can achieve large-scale deployment in homes and industrial settings.

Financing momentum in both domestic and international primary markets continues to surge. Ant Lingbo has completed a 1.5 billion yuan independent funding round, while overseas players Skild AI and Physical Intelligence have secured billion-dollar-level investments in succession. Capital is now independently assigning valuations to robot brains, signaling a critical window for the re-rating of the physical AI industry chain. The entire ecosystem has formed a closed data loop spanning upstream perception and data collection hardware, midstream simulation foundations and physical foundation models, and downstream real-world robot deployment—unlocking investment opportunities across the full chain from hardware perception and simulation training to iterative physical testing.

Upstream perception and data collection hardware serves as the gateway for physical AI to acquire real-world ground truth data, supplying midstream simulation models with multimodal raw training material. Demand for tactile sensing, force sensing, and edge-side computing hardware is erupting simultaneously. Fulaisi New Material, a core player in electronic skin technology—often described as the robot's "third brain"—has pioneered its integrated sensing and control tactile intelligent TPU, with order backlog reaching one million units. The company has partnered with Haocun Technology to jointly develop data collection gloves, addressing the shortfall in robotic tactile perception. Tactile information is a critical input for robots executing fine manipulation tasks, and electronic skin products directly serve the real-world data collection segment, positioning the company to benefit deeply from the explosive demand for multimodal training data in physical AI.

Keli Sensing focuses on six-axis force sensors, building a multimodal perception system that layers force sensing atop tactile sensing. These sensors are core hardware for embodied intelligence real-world data collection. When robots perform contact-based tasks, force feedback is indispensable, and the company's products are widely used in real-machine data acquisition, providing force-dimensional ground truth data for physical AI models. Computing hardware, responsible for local preprocessing of multimodal data, constitutes an essential complement in the data collection chain. Rockchip has developed proprietary NPU chips widely deployed in first-person-view (EGO) data collection devices and robot perception boxes. These chips handle local preprocessing of multi-camera and IMU multimodal data, completing initial data processing at the edge and reducing computational pressure on backend simulation platforms, deeply embedding the company within the physical AI data collection chain. SigmaStar Technology's edge-side vision SoCs integrate visual perception with edge computing capabilities, supporting various visual data collection terminals and providing computational support for first-person-view robot data acquisition, with edge-side visual information processing adapted to virtual-real integrated training modes.

The midstream simulation foundation and physical AI models represent the segment with the highest technical barriers in the entire industry chain. These platforms leverage real-world collected data upstream to construct virtual physical environments, generate synthetic training data at scale, and output intelligent policies directly deployable on physical robots. Suocheng Technology has independently developed the TianGong series of physical solvers, creating virtual training grounds for humanoid robots capable of multi-field coupled simulation including rigid bodies, soft bodies, and fluids. Benchmarking against leading overseas simulation platforms, the company provides virtual training environments for physical AI foundation models. These virtual training grounds dramatically reduce the cost of physical robot training while enabling rapid iteration of robotic action policies, positioning the company as a scarce domestic underlying simulation engine play.

Nengke Technology has developed the Ling series of industrial embodied intelligence agents, training physical AI models using real-world industrial sensor data. Focusing on robotic arms, inspection robots, and other industrial scenarios, the company has established a complete closed-loop pathway from simulation training to industrial physical robot iteration, positioning it to capture near-term benefits from industrial embodied intelligence deployment. Ant Lingbo, as an industry benchmark player, has independently developed the LingBot world model and VLA embodied foundation models, targeting the general-purpose robot brain market and accelerating the maturation of China's simulation and physical world model ecosystem.

Downstream, humanoid robots and industrial robotic arms serve as physical carriers that complete actual operations while continuously generating new perception data that feeds back upstream, fulfilling the closed loop of "real-world collection, simulation training, and iterative physical deployment." Estun Automation, as a domestic leader in industrial robots, imports real-world operational data—including assembly parameters, contact forces, and motion trajectories—into simulation platforms for robot policy pre-training and virtual-real validation. As its hardware business scales, the company continuously produces high-quality industrial datasets that feed back into physical AI model iteration. Dobot Technology's robotic arms incorporate force control and visual perception modules, collecting tactile, force, and visual multidimensional data during precision interaction tasks, building industrial interaction datasets that supply high-quality industrial scenario training materials to midstream physical AI models. Unitree Robotics and Zhiyuan Robotics, as representative humanoid hardware companies, have established real-world data collection factories that capture multimodal data and achieve virtual-real closed-loop iteration, driving sustained demand for upstream perception hardware and simulation software.

Taken together, the physical AI investment thesis can be mapped along the closed-loop chain of data generation, model training, and physical deployment. Upstream, priority should be given to tactile and force sensing hardware along with edge-side computing chips—data remains the foundation for physical AI iteration, and data collection hardware will be the first to experience demand expansion. Midstream, attention should focus on domestic simulation engines and industrial physical AI platform companies; simulation foundations represent essential underlying industry tools with high technical barriers and long-term growth potential. Downstream, the focus is on robot hardware leaders capable of continuously producing high-quality datasets, leveraging real-world scenarios to complete the data loop while simultaneously driving adoption of upstream software and hardware products. As primary market capital intensifies its commitment to the robot brain track, the industry is transitioning from hardware-centric competition to intelligent systems competition, with the physical AI foundation model industry chain entering a phase of industry-capital resonance and synchronized growth.