For a humanoid robot, trying to sink a basketball shot is not as simple as it seems to a human athlete. It requires countless hours of fine-tunings and calculations by engineers before the mechanical arm can drop the ball through a hoop.

Fortunately, the wide gap between virtual simulation and real-world execution is closing fast, driven by the advancement of homegrown technology. This breakthrough arrives as China's 15th Five-Year Plan (2026-2030) prioritizes foundational AI research and technology transfer.

First differentiable physics engine

In this case, the innovation comes from Fysics, China's first differentiable physics engine, developed by Fysics AI in collaboration with Fudan University.

Unlike traditional physics engines that only simulate forward motion, Fysics is differentiable — meaning it can calculate errors and backpropagate corrections.

"Traditional physics engines are like one-way streets," said Zhang Lihua, former key contributor to NVIDIA's PhysX engine and now head of the team behind Fysics. "They can simulate movement but cannot tell you where the errors came from."

If a robot misses a basketball shot, Fysics analyzes whether the force was too strong, the angle was off, or the timing wrong. The robot then adjusts autonomously, without tedious real-world trial and error.

Solving the 'sim-to-real' bottleneck

Zhang calls physical AI "the only path for AI to move from the virtual to the real world."

Fysics combines a unified multi-material solving framework with high-precision contact resolution. This allows robots to learn precise manipulation with minimal trial and error, overcoming the sim-to-real transfer challenge that has long plagued embodied AI, humanoid robotics, autonomous driving, and industrial digital twins.

Building an ecosystem

Fysics is not a standalone entity. Zhang's team has built a complete infrastructure including MoziSim, a simulation platform for generating massive, realistic training data, and OmniFysics, a multimodal foundation model for perceiving and reasoning about physical reality.

In the next three to five years, the team aims to create a complete industrial chain combining domestic computing power, proprietary engines, and open-source ecosystems. Fysics has already partnered with Chinese chipmakers and research institutions to scale the technology.

"We want to become a core infrastructure provider for the age of physical intelligence," Zhang said.

For now, the basketball-shooting robot is proof that Chinese innovation is helping AI learn to not only think, but also to act and adapt.