跳转到主要内容

Unitree Accelerates Self-Evolution of Physical AI Models, Wang Xingxing Reveals New Research Directions

At the main forum of the 2026 World Robot Conference, Wang Xingxing, founder and chairman of Unitree, brought exciting news: the company is advancing the self-evolution of physical AI robot models and attempting to use large AI models to automatically complete the development of robot control code. Just one day after Unitree's listing on the STAR Market, Wang Xingxing shared this cutting-edge research direction.

Wang Xingxing stated that the company plans to use cutting-edge AI large models as the core, customize rules, experience frameworks, and constraint tools, allowing the models to autonomously search for the latest academic papers, research results, and open-source solutions, generate robot control code, and then verify through model evaluation and manual review. This closed-loop system not only improves development efficiency but also incorporates simulation data, real-world data, and human behavior data into the training process. As the scale of robot deployment expands, the accumulation of test data and evaluation metrics is expected to significantly accelerate robot development and iteration.

At the conference, Unitree showcased the evolution of robots from single-agent intelligence to swarm intelligence, and from motion control to autonomous collaboration. Multiple humanoid robots, quadruped robots, and wheel-legged robots completed cluster collaboration based on the company's self-developed AI swarm control system. Additionally, the manned exoskeleton GD01 made its debut.

Regarding the 'ChatGPT moment' for embodied intelligence, Wang Xingxing believes that the biggest bottleneck remains insufficient generalization capabilities. He suggested that when robots can complete about 80% of tasks in unfamiliar environments through voice or language instructions alone, the industry may see breakthrough progress. This time point might be within 2 to 3 years, or it could be delayed to 5 to 10 years.

Wang Xingxing also pointed out that current robot models struggle to effectively correct tactile errors at the centimeter or millimeter level in fine operations, with the root cause being the discrepancy between AI model inputs/outputs and the real world. With continuous technological breakthroughs, such cumulative errors are expected to improve in the coming years.

Key Points

  • Unitree is advancing the self-evolution of physical AI models, using large models to automatically generate control code.
  • The closed-loop system integrates simulation, real, and human behavior data to improve development efficiency.
  • Multi-robot collaboration and the manned exoskeleton GD01 were showcased at the conference.
  • Wang Xingxing predicts that the 'ChatGPT moment' for embodied intelligence may arrive within 2 to 3 years.
  • The current main bottlenecks are insufficient generalization capabilities and tactile error issues.