Freshly Listed on STAR Market, Unitree's Wang Xingxing Speaks at WRC: How Humanoid Robots Transition from 'Showpieces' to 'Products'

The day after Unitree Technology (688836-CN), dubbed "China's first humanoid robot IPO," listed on the Shanghai Stock Exchange's STAR Market, its founder, chairman, and chief technology officer, Wang Xingxing, attended the main forum of the 2026 World Robot Conference (WRC). Wang delivered a keynote address titled "From Showpiece to Product: The Next Decade of the Humanoid Robot Industry."

During his speech, he shared personal insights and Unitree’s latest pre-research directions regarding the practical implementation of embodied intelligence, the industrial inflection point for explosive growth, and the technical pathways toward robotic self-evolution.

### Predicting the 'ChatGPT Moment' for Embodied Intelligence: The Rigorous 80-80 Criterion

On the widely discussed question of when general-purpose robots will enter everyday life and homes, Wang offered a cautious and objective forecast. He believes that for embodied intelligence to reach an industry-wide inflection point akin to the "ChatGPT moment," it may take as soon as two to three years, or as long as five to ten years.

Wang emphasized that this inflection point should not be equated with technical demonstrations in specific scenarios. Instead, he proposed a stricter evaluation standard: In 80% of unfamiliar environments, robots must autonomously complete approximately 80% of tasks using only voice or text instructions. He noted that the use of the word "toward" in his presentation slides signified that this threshold has yet to be crossed—a technical barrier requiring collective breakthroughs across the industry.

### Core Industry Bottleneck: The 'Final Few Centimeters' of Fine Manipulation

Regarding generalization capability, Wang candidly stated that "insufficient generalization ability" is currently the biggest bottleneck faced globally. In fixed environments with sufficient training data, a robot’s task success rate can approach 100%. However, even slight changes in the object being manipulated or the environment cause a significant drop in success rates.

He analyzed that the specific bottleneck lies in the micro-manipulation phase—the "final few centimeters" or even "final few millimeters"—of task execution. At this stage, while the robot’s high-level understanding and task decomposition are sound, minor tactile feedback discrepancies or positional deviations cannot be corrected in real time, leading to cumulative errors that ultimately result in failed grasps or dropped objects.

Wang pointed out that this stems from a fundamental issue: the misalignment between AI model inputs/outputs and actual physical robots. Unlike language models whose inputs and outputs exist entirely within lossless digital encoding and vector spaces, every input and output of a physical robot interacting with the physical world inevitably involves deviation and loss. Nevertheless, he expressed optimism that this problem could be resolved within the next few years.

### In-House Technological Exploration: The 'Physical AI Robot Self-Evolution' Framework

To improve development efficiency, Unitree unveiled a pre-research framework labeled "Physical AI Robot Self-Evolution V1.0." This technical roadmap envisions directly leveraging cutting-edge large AI models to drive computation and evolution in physical-world robots.

The framework operates as follows: Engineers define rules, experiential constraints, and tool interfaces. An AI large model then autonomously searches the web for the latest academic papers, high-quality research, and open-source solutions, and independently generates control code for robots. After code generation, it is first validated and trained in a simulation environment, then deployed onto physical robots for testing. Test results are jointly evaluated and scored by the AI model and human assessors, with feedback looped back to the programming agent, forming a positive self-cycling closed loop.

Wang stated that this self-evolution system can effectively adapt to continuous improvements in foundational models, enhance the utilization of multi-source data, and enable skills to accumulate and solidify daily or monthly as the scale of physical robot deployment expands. However, he stressed that this technology remains in the company’s pre-research phase and is not yet a commercially validated outcome.

In closing, Wang noted that Unitree has participated continuously in the World Robot Conference since 2017 or 2018. At this juncture of AI explosion and global attention, Physical AI is reaching a brand-new starting point, and the pace of robotic evolution and development may accelerate faster than anticipated.