Shihang Intelligence's ocean general-purpose foundation model "Cangqiong CEORION" completes filing with the Cyberspace Administration of China

The ocean general-purpose foundation model "Cangqiong CEORION," independently developed by Shihang Intelligence, recently completed official filing with the Cyberspace Administration of China. This is the world's first ocean general-purpose foundation model. The model is based on Shihang's original LWAM unified architecture (Language-World-Action Model), dynamically fuses multi-dimensional sensor data such as vision, acoustics, pressure, fluid dynamics, and attitude, and builds an end-to-end unified neural network capable of integrating semantic understanding, environmental modeling, and action generation, while achieving a fully autonomous closed loop covering underwater environmental perception, task assessment, path planning, autonomous operation, and emergency response, as well as general intelligent capabilities with exponential generalization. This progress marks the first time a Chinese enterprise has, in the global frontier field of ocean general-purpose artificial intelligence, successfully established the complete chain from underlying technology self-development and model training iteration to compliant industrialized services. In this field, China has no ready-made international paradigm or experience to follow, and has explored a path of independent innovation. The marine environment is highly complex. As operating depth increases, the pressure borne by underwater equipment continues to rise; complex ocean currents, surges, and water disturbances affect the stability of robot movement and operation. Underwater lighting is insufficient, and turbid water, suspended matter, and complex backgrounds further increase the difficulty of environmental perception and target recognition, making it difficult for models that rely heavily on vision to truly function. In addition, underwater communication conditions place higher demands on autonomous intelligence. Long-distance underwater communication usually faces limitations such as limited bandwidth, high latency, and unstable connections, making it difficult to transmit data back to the cloud for processing and difficult to continuously rely on high-bandwidth links for remote control. This requires robots, under conditions of incomplete information and limited communication, to still be able to combine on-site perception for task planning, risk judgment, and autonomous execution. A deeper challenge lies in data and generalization. Marine operation data collection is costly and time-consuming, and there are significant differences among different sea areas, depths, water quality, lighting conditions, and robot platforms. Models trained on data from specific scenarios may show obvious performance degradation when the environment or task changes. Therefore, what the ocean general-purpose foundation model needs to solve is not only recognizing marine targets or understanding marine knowledge, but also how to effectively combine environmental perception, task understanding, action planning, and robot execution, and to explore the model's generalization ability across unseen environments, tasks, and equipment platforms. This is also an important technological foundation for marine intelligence to move from single-function automation to a higher level of autonomous intelligent operation. To make breakthroughs in these key technological innovations, Shihang has relied on 19 years of sustained technological accumulation and massive real-world data, as well as the persistent belief in daring to be first in the world. The ocean general-purpose foundation model "Cangqiong CEORION" has delivered an answer sheet with multiple world firsts and world premieres: In elastic modal perception technology, the brain automatically adjusts perception channels and intelligently fuses sonar, force perception, inertial navigation, vision, and more, achieving stable perception and operation in extreme waters with zero illumination and high turbidity. In the native ocean world model, for the first time, complex marine fluid dynamics and kinematic constraints are encoded into the underlying layer of the neural network, enabling precise prediction of ocean current impacts and force feedback in the ocean, and achieving active real-time adaptive control. In terms of cross-scenario and cross-task generalization and universality, Shihang, centering on real scenarios of complex marine environments and diverse operations, pioneered marine expert models in multiple fields and, through the LWAM unified architecture, achieved collaboration and transfer of professional capabilities, driving the model from specialized capabilities for specific tasks toward cross-scenario and cross-task generalization and autonomous execution, and achieving exponential generalization and general capabilities across all scenarios. In terms of the connection between the model and the body and adaptation to diverse scenarios, Shihang's Cangqiong large model, based on a unified end-to-end architecture, can support 12 major categories of underwater operations—including inspection, detection, cleaning, grasping, cutting, welding, exploration, search and rescue, emergency response, mining, operation and maintenance, and storage—without switching between different models, driving marine intelligence from single-task execution toward multi-task, multi-scenario general autonomous operation. The value of technological breakthroughs must ultimately be verified in real industries. From technological innovation to practical operations, from single-point breakthroughs to multi-scenario expansion, Shihang is continuously driving ocean general-purpose intelligent capabilities toward industrial application. Relying on abundant real operation scenarios, Shihang continuously accumulates high-value data, iteratively optimizes models, and uses improved model capabilities to expand more application scenarios, forming a growth flywheel in which "scenarios—data—models—applications" reinforce one another. At present, the relevant capabilities have been implemented in fields such as ship cleaning, port dredging and green operations, aquaculture, marine engineering and infrastructure, and deep-sea resource exploration and development, and have gradually achieved commercial deployment and large-scale application, demonstrating the ability to transform from underlying technological innovation to industrial value creation.