Advancing AI: How Physics Engines Bridge the Gap Between Simulation and Reality

In the demonstration hall of Feijie Kesi Intelligent Technology (Shanghai) Co., Ltd., a video plays on a large screen, showcasing a seemingly simple yet profoundly significant scene: a robot repeatedly attempts to shoot a basketball into a hoop. Initially lacking direction, it gradually adjusts its posture and strength, eventually succeeding in making the shot.

For human basketball players, this task is effortless; however, for robots, it was once a formidable barrier between virtual and real worlds.

The backbone that enables robots to master the “shooting” skill is China’s first differentiable physics simulation engine, Fysics. The development of this engine marks a significant breakthrough in the core technology of physics AI, allowing China to secure a vital position in the global competition.

The “14th Five-Year Plan” clearly emphasizes the strengthening of technological layouts in strategic frontier areas, advocating for the implementation of artificial intelligence and other technological strategies, accelerating breakthroughs in foundational theories and core technologies, and promoting their transformation and application.

“The second phase of AI is a competition regarding the understanding of the physical world,” stated Zhang Lihua, Vice Dean of the School of Intelligent Robotics and Advanced Manufacturing Innovation at Fudan University and founder of Feijie Kesi. He added that in the next 3 to 5 years, Feijie Kesi aims to create a complete chain of “domestic computing power + autonomous engine + open-source ecosystem” to accelerate the industrialization of physics AI.

Core Breakthrough: Upgrading from “Only Calculating Results” to “Self-Correction”

Once AI learns to see, write, and generate, what is the next step? The global tech community unanimously agrees: physics AI—enabling intelligent agents to truly understand the rules of the real world, such as gravity, friction, collisions, and deformation, thus achieving autonomous reasoning, stable interactions, and reliable decision-making.

“Physics AI is an essential path for artificial intelligence to transition from the virtual realm to reality, and from perception to free interaction,” Zhang Lihua told reporters from the Science and Technology Daily. “It has become a strategic high ground in the global AI industry competition.”

At the International Consumer Electronics Show held this January, Nvidia CEO Jensen Huang mentioned the concept of physics AI 17 times during his keynote speech. “From humanoid robots and embodied intelligence to industrial digital twins and autonomous driving, the intelligent upgrades of these trillion-dollar industries urgently require a foundational platform for deep integration with AI,” Zhang Lihua noted.

As one of the primary founders of Nvidia’s PhysX physics engine and the world’s first real-time commercial physics simulation solution, Zhang Lihua is determined to help China achieve autonomy and control in this crucial technology. “Traditional physics engines lack differentiability and sufficient simulation accuracy, which have long constrained industry development. There are very few companies globally with the capability to independently develop differentiable physics simulation engines,” he explained. “The domestic market has long relied on foreign physics engines and simulation platforms, making the demand for autonomy urgent.”

Building on previous technological achievements, Zhang Lihua’s team consolidated research resources from schools and enterprises to successfully develop Fysics, China’s first differentiable physics engine.

What does “differentiable” mean? Zhang Lihua explains that traditional physics engines operate like a one-way street, only able to simulate object motion in one direction without being able to feedback error sources. In contrast, a differentiable physics engine creates a two-way channel, allowing for gradient backpropagation to directly inform the system where it went wrong and how to correct it.

Using the example of the robot shooting a basketball, if it misses, Fysics can utilize its differentiable capabilities to indicate whether the issue was too much force or an incorrect angle, enabling the robot to autonomously adjust its strategy. Coupled with a carefully designed unified solution framework for multiple physical materials and high-precision contact resolution, Fysics allows the robot to learn precise operations without extensive trial and error, achieving a seamless transition from simulation to reality and overcoming one of the most challenging hurdles in the realm of embodied intelligence.

“We have upgraded the physics engine from ‘only calculating results’ to ‘self-correction’, bridging the final mile for AI from simulation to reality,” Zhang Lihua stated.

Industry Layout: Promoting the Transition of Physics AI from Laboratory to Large-Scale Implementation

Zhang Lihua’s team aims to construct an operating system for the era of physics AI. On top of the Fysics engine, which serves as the “foundation,” Feijie Kesi has also developed the MoziSim simulation training platform as its “factory” and the OmniFysics multimodal physics AI foundational model as its “brain.” “The differentiable physics engine defines the computational rules of the physical world; the simulation platform achieves large-scale data production capabilities by creating a realistic digital world; and the foundational model enables AI to perceive, comprehend, and reason about the physical world,” Zhang Lihua explained, likening this to creating a universal operating system for intelligent agents in the real world.

The team also released the world’s first comprehensive benchmark for physical perception and logical reasoning assessment, aiming to define the standards for physics AI.

Industry experts believe that this entire stack of physics AI, from engines and platforms to models and assessment standards, will accelerate the industrialization of physics AI in China. In the realms of humanoid robots and embodied intelligence, it will significantly reduce training costs, enhance the gait stability, precise operation, and environmental adaptability of general-purpose robots, thereby facilitating their integration into factories and homes. In the area of industrial digital twins, it can simulate equipment operation, material interaction, and production line processes with high precision, supporting production line optimization and predictive maintenance, significantly reducing costs and increasing efficiency.

Zhang Lihua stated that Feijie Kesi is collaborating with domestic chip manufacturers, research institutions, and other stakeholders to build an industrial ecosystem. Previously, Feijie Kesi has engaged in deep cooperation with several ecosystem partners, including Muxi Co., Parallel Technology, Tuosida, and Lijie Technology, focusing on computing power adaptation, scene implementation, and ecological collaboration.

“In the next 3 to 5 years, we hope to achieve product maturity across the stack and domestic adaptation, becoming a core infrastructure provider for the global era of physics intelligence and forming a complete industrial chain of ‘domestic computing power + autonomous engine + open-source ecosystem’ to promote leapfrog development in the fields of physics intelligence and embodied intelligence in China,” Zhang Lihua concluded.