At the 2026 World Robot Conference Forum, held in Beijing from August 19 to 22, Wang Xingxing, founder and chairman of Unitree Technology Co., Ltd., delivered a keynote speech outlining a groundbreaking initiative his company introduced just yesterday and is now actively implementing. He highlighted that while AI has been widely adopted for programming and development tasks in recent years, its practical application within the robotics sector remains notably limited.

To address this gap, his team has been driving a key initiative focused on enabling the autonomous evolution of physical AI robot models. By leveraging cutting-edge large AI models, they have established their own rules, accumulated experiential data, and defined constraints and tool systems. This framework empowers robots to independently search the internet for the latest research papers, high-quality findings, and open-source solutions. From there, the robots autonomously program and generate proprietary control codes tailored to their operations.

Wang elaborated that once the control code is generated, the robot can run it in a simulated environment, utilizing large model capabilities to subsequently manage the physical robot. The team then proceeds with deployment testing on the actual hardware, followed by evaluation and scoring. This assessment process comprises two components: one is conducted autonomously by the robot's AI model, while the other involves human evaluators who provide their own ratings.

"Human experience is essential, as people can intuitively assess the effectiveness of outcomes, and this is critical to the process," Wang noted. "Once this logic system operates smoothly, the overall efficiency of robot development will surge, and iteration speeds will accelerate dramatically. After running this workflow for a period, we can genuinely achieve a leap in the robots' self-evolution capabilities," he added.