Wang Xingxing, founder and chairman of Unitree Robotics, revealed at the main forum of the 2026 World Robot Conference that the company is advancing the self-evolution of physical AI robot models and exploring the use of large AI models to automatically complete robot control code development. On the day after Unitree Robotics listed on the Sci-Tech Innovation Board, Wang Xingxing stated that the company plans to take cutting-edge AI large models as the core, customizing rules, experience frameworks, and constraint tools, allowing the model to autonomously search for the latest academic papers, research results, and open-source solutions, generate robot control code, and then verify it through model evaluation and manual review.
Wang Xingxing pointed out that improvements in basic model capabilities will drive the continuous iteration of the self-evolution system. This closed-loop system can also incorporate simulation data, real-world data, and human behavior data into the training process. As the deployment scale of robots expands, test data and evaluation metrics continue to accumulate, which is expected to significantly improve the efficiency of robot development and iteration.
At the conference, Unitree Robotics demonstrated the evolution of robots from single intelligence to group intelligence, and from motion control to autonomous collaboration, through multi-machine coordination. Multiple humanoid, quadrupedal, and wheel-legged robots completed cluster collaboration based on the company's self-developed AI group control system, and the manned exoskeleton GD01 was also unveiled simultaneously.
Regarding the "ChatGPT moment" of embodied intelligence, Wang Xingxing believes the biggest bottleneck remains insufficient generalization ability. When robots can complete about 80% of tasks by only voice or language commands in unfamiliar environments, the industry may reach a breakthrough threshold, which could happen in 2 to 3 years, or even up to 5 to 10 years if delayed.
He believes that current robot models struggle to effectively correct tactile errors in fine operations within a few centimeters or millimeters, with the root cause being the discrepancy between AI model inputs/outputs and the real world. With continued technological breakthroughs, such accumulated errors are expected to be improved in the coming years.
