In the demonstration hall of Feijie Kesi Intelligent Technology (Shanghai) Co., Ltd., a seemingly simple but meaningful video is scrolling on the large screen: a robot continuously tries to throw a basketball into the basket, starting from shooting randomly without any sense of direction, gradually adjusting its posture and strength, and finally steadily sending the basketball into the basket. For human basketball players, this is a small effort; For robots, this used to be a chasm between virtual and reality. The hardcore support for robots to master the "shooting" skill is Fysics, the first differentiable physics simulation engine in China. The birth of this engine marks China's independent breakthrough in the underlying core technology of physical AI and occupies an important position in global competition. The outline of the 15th Five Year Plan clearly strengthens the layout of science and technology in strategic frontier fields, proposes the implementation of strategic deployments such as artificial intelligence, accelerates breakthroughs in basic theories and underlying technologies, and promotes transformation and application. The second half of AI is a competition about understanding the physical world. ”Zhang Lihua, vice president of the Institute of Intelligent Robotics and Advanced Manufacturing Innovation of Fudan University and founder of Feijie Kesi, said that in the next 3-5 years, Feijie Kesi will build a complete chain of "domestic computing power+independent engine+open source ecology", and promote the acceleration of physical AI towards industrialization. Core breakthrough: Upgrading from "only counting results" to "autonomous error correction". After AI learns to read, write, and generate, what is the next step? The unanimous answer given by the global technology community is: physical AI - enabling intelligent agents to truly understand real-world rules such as gravity, friction, collision, deformation, etc., enabling autonomous reasoning, stable interaction, and reliable decision-making. Physical AI is a necessary path for artificial intelligence to move from virtual to reality, from perception to free interaction, "Zhang Lihua told Science and Technology Daily reporters." It has become a strategic high ground for global AI industry competition. ”At the International Consumer Electronics Show held in January this year, Nvidia CEO Huang Renxun mentioned the concept of physical AI 17 times in his keynote speech. From humanoid robots and embodied intelligence to industrial digital twins and autonomous driving, the intelligent upgrading of trillion level physical industries urgently requires a physical simulation base that can deeply integrate with AI. ”Zhang Lihua said. As one of the main founders of NVIDIA's PhysX physics engine and the world's first commercial real-time physics simulation solution, Zhang Lihua is determined to promote China's independent and controllable development in this key core technology. Traditional physics engines are non differentiable and have insufficient simulation accuracy, which are the core pain points that have long constrained the development of the industry. There are only a few companies worldwide that have the ability to independently develop differentiable physics simulation engines, "he said." The domestic market has long relied on foreign physics engines and simulation platforms, and there is an urgent need for independent controllability. "Based on previous technological accumulation, Zhang Lihua's team integrated the scientific research capabilities of schools and enterprises and successfully developed the first differentiable physics engine Fysics in China. What is' differentiable '? Zhang Lihua explained that traditional physics engines are like a one-way street that can only simulate object motion in a forward direction and cannot provide feedback on error sources; And the differentiable physics engine builds a bidirectional channel that can directly tell the system where something is wrong and how to correct it through gradient backpropagation. Zhang Lihua takes robot shooting as an example: if it doesn't make a shot, Fysics will use its differentiable ability to tell the robot whether the force is too strong or the angle is off, allowing it to adjust its strategy autonomously. In addition, with a carefully designed unified solution framework for multiple physical materials and high-precision contact calculation, Fynics enables robots to learn precise operations without massive trial and error, truly achieving seamless transfer between simulation and reality, and overcoming the most difficult problem in the field of embodied intelligence from simulation to reality. We have upgraded the physics engine from 'only calculating results' to' autonomous error correction ', bridging the last mile of AI from simulation to reality. ”Zhang Lihua said. Industrial layout: Promoting the large-scale implementation of physical AI from the laboratory. The goal of Zhang Lihua's team is to build an operating system for the era of physical AI. On the foundation of the Fytics engine, Feijie Kesi has also built the MoziSim simulation training platform as a "factory" and the OmniFytics full modal physics AI basic model as a "brain". The differentiable physics engine is the core foundation that defines the computational rules of the physical world; the simulation platform achieves large-scale data production capabilities by building a real digital world; and the basic model enables AI to perceive, understand, and reason about the physical world. ”Zhang Lihua said that this is equivalent to creating a universal operating system for real-world intelligent agents. The team has also released the world's first comprehensive physical perception and logical reasoning evaluation benchmark, aiming to define the standards for physical AI. The industry believes that the entire physical AI stack foundation, from the engine, platform, model to evaluation standards, from the bottom to the application, will accelerate the industrialization of physical AI in China. In the field of humanoid robots and embodied intelligence, it will significantly reduce training costs, improve the gait stability, fine operation, environmental adaptation and other abilities of general robots, and accelerate the entry of general robots into factories and homes; In the field of industrial digital twins, high-precision simulation of equipment operation, material interaction, and production line processes can be achieved, supporting production line optimization and predictive maintenance, significantly reducing costs and increasing efficiency. Zhang Lihua said that Feijie Cosis is working with domestic chip manufacturers, scientific research institutions and other forces to jointly build an industrial ecology. Previously, Feijie Kesi had deepened cooperation with multiple ecological partners such as Muxi Shares, Parallel Technology, Tuosida, and Lijie Technology in areas such as computing power adaptation, scene implementation, and ecological collaboration. In the next 3 to 5 years, we hope to achieve full stack product maturity and domestic adaptation, become a core infrastructure provider in the global era of physical intelligence, form a complete industrial chain of 'domestic computing power+independent engine+open source ecosystem', and promote China's leapfrog development in the fields of physical intelligence and embodied intelligence. ”Zhang Lihua said. (New Society)