Wang Xingxing Sets 'Two 80%' Benchmark: Quantitative Marker for Robot Explosion Point
On August 19, 2026, Unitree Robotics listed on the STAR Market, with its stock surging over 600% on the first day, briefly pushing its market value above ¥440 billion. The next day, founder Wang Xingxing appeared at the 2026 World Robot Congress in Beijing, delivering a speech titled "From Exhibit to Product: The Next Decade of Humanoid Robots," focusing entirely on technology rather than stock prices.
In his speech, Wang proposed a quantitative threshold for the explosion of embodied intelligence: when a robot is placed in 80% unfamiliar environments and can successfully complete about 80% of tasks based on voice or text commands, that marks the "ChatGPT moment" for embodied AI. He estimated this could occur in as little as two to three years, or up to five to ten years at the latest.
This benchmark converts the previously conceptual "generalization ability" into a measurable goal. At the 2026 World Robot Congress, generalization was a buzzword among executives. Some analysts noted that capital markets no longer directly pay for generalization imagination but instead price it indirectly through four proxy variables: shipment volume, order quality, convergence of technical routes, and data flywheel.
Wang explained a key technical difficulty: why robots can achieve near-100% success in fixed scenarios after adequate training, but performance plummets when environments or objects change slightly. The issue lies in the misalignment between AI models and the real physical world. While language models operate in lossless digital vector spaces, each physical action introduces errors that accumulate over several steps — a deviation of a few millimeters can cause a grasping task to fail entirely.
This fundamental difference distinguishes embodied intelligence from traditional large models. Digital models process information purely, whereas embodied robots continuously interact with the physical world, generating errors with every interaction.
To address this challenge, Wang disclosed a pre-research project called "Embodied AI Robot Self-Evolution V1.0." The core logic is to let AI take over robot development itself: AI automatically searches the latest papers and open-source solutions, writes control code, validates in simulation, deploys to physical robots for testing, and then both AI and humans evaluate and score, creating a closed-loop iteration.
He believes this mechanism offers triple advantages: continuous improvements in foundation models will boost the self-evolution loop; it efficiently utilizes diverse data, breaking through manual collection bottlenecks; and as robot deployment expands, data utilization and skill accumulation will generate economies of scale. Wang's judgment is clear: the industry's biggest bottleneck is insufficient generalization, and the breakthrough lies in aligning AI model inputs and outputs with the real physical world, a problem he believes will be solved within a few years.