Chinese autonomous driving company QCraft unveiled its physical AI model and QPilot MAX assisted-driving solution at Auto China 2026 in Beijing on Friday, as it seeks to apply physical AI technologies to assisted-driving systems and Level 4 applications such as robotaxis and autonomous logistics vehicles.

The company said the physical AI model is built on a framework combining world models and reinforcement learning. The model includes a cloud-side component used offline to generate training data and simulate complex traffic scenarios, while the in-vehicle model handles driving tasks.

QCraft said QPilot MAX is an assisted-driving solution supported by more than 500 TOPS of computing power and is designed to improve performance in complex urban traffic conditions.

The company said its broader assisted-driving systems are available on 25 production models and are expected to be added to more than 50 models in 2026. The scale of deployment gives QCraft more real-world data and helps validate system reliability, it said.

Yu Qian, co-founder and CEO of QCraft, said autonomous driving is one of the clearest early applications for physical AI. He said physical AI remains at an early stage and still needs further technological advances before wider adoption. Autonomous driving, however, already has large volumes of driving data and a more mature engineering base, making it a practical starting point, he added.

Yu said QCraft is using world models and reinforcement learning to improve the training of autonomous-driving systems, as real-world testing is time-consuming and may not cover enough rare or complex traffic scenarios.

The approach allows the company to generate more scenarios in simulation and use the resulting data to improve in-vehicle driving systems, he said.

Yu said QCraft is focusing on improving the vehicle's AI decision-making capability rather than simply adding more sensors or computing power.

The company also outlined its Level 4 robotaxi and autonomous logistics vehicle programs. Yu said robotaxi commercialization will require mass production and large-scale validation, adding that QCraft's current priority is to improve the core driving capability of its AI system.

Yu said QCraft's assisted-driving systems and Level 4 applications draw on similar underlying model capabilities, but Level 4 vehicles require additional sensors, computing capacity and redundant safety systems.

Li Dong, CTO of QCraft, said the key issue for Level 4 deployment is whether model capability can support safe operations at a commercially viable cost. As driving models become stronger, Level 4 services could cover more scenarios and operate more efficiently, Li said.

QCraft said it will continue to develop both mass-produced assisted-driving system and Level 4 applications, including robotaxis and autonomous logistics vehicles.