Unitree Founder Wang Xingxing: Robot Boom Awaits Two '80%' Moments

Wang made the remarks at the main forum of WRC 2026 in Beijing, according to a report by Leiphone. He said the biggest bottleneck for embodied AI is not whether a robot can perform a single action, but whether it can generalize to a different room, object, or task. While demonstration videos may show impressive feats such as running at 10 meters per second, real-world adoption requires working in new settings without retraining for every task.

He also introduced his idea of "self-evolution" for physical AI robots, which he first proposed at the company's IPO celebration banquet a day earlier. The plan is to use advanced foundation models to let AI search for research papers, generate control code, run simulations, conduct tests on physical robots, and score the results, with feedback feeding back into the loop. This would allow the robot development process to increasingly run on its own.

In his speech, Wang gave several concrete examples of recent progress. In April, a modified version of Unitree's H1 humanoid robot set a world record with a peak running speed of about 10 meters per second, or 36 km/h. In May, the company released what it called the world's first mass-produced manned mech, a roughly three-meter-tall vehicle weighing around 500 kilograms when occupied, capable of traversing rough outdoor terrain and transforming into a four-legged mode for better stability.

He also discussed the company's work on end-to-end, multimodal AI that can generate robot movements in real time based on voice commands. Currently, there is a delay of several seconds because the system performs speech recognition, cloud-based action generation, and validation before sending commands to the robot. Wang acknowledged that the generated actions are not yet perfectly consistent; saying the same sentence on different days may lead to slightly different movements.

The company has also been training robots in its own meeting rooms to clean up cluttered spaces, a task involving seven to eight different sub-tasks controlled by a single model with some anti-interference capability. Wang called this kind of relatively practical but not overly difficult scenario meaningful for getting robots into workplaces.

He highlighted the recent launch of the Unitree As2-W, a lightweight wheeled quadruped robot weighing about 25 kilograms, with rain resistance and strong load and endurance capabilities for both indoor and outdoor use. The company also previewed a humanoid robot nicknamed "Superman," developed in just over three months, which has hit a top speed of 12.66 meters per second and a jumping height exceeding the best human performances in history. Wang stressed that the preview is not a formal product release.

Turning to AI models, Wang said the company has been investing heavily in video-generation-based world models, which he said have become a popular direction this year. Unitree first began working on such models in early 2020 but paused because results were not satisfactory, then resumed last year.

He attributed the current gap in generalization to a mismatch between AI models' inputs and outputs and the physical world. Language models operate on digitally encoded vectors and do not generate physical errors, but each input-output cycle of a robot can introduce deviations. These errors often appear in the last few millimeters when a robot tries to grasp an object, causing the whole task to fail.

He estimated that reaching the two 80% thresholds could take two to three years in the best case, or five to ten years in the worst case. He also suggested that robot self-evolution could benefit from using more diverse data, increasing real-robot deployment, accumulating skills over time, and following the progress of foundation models. As human engineers increasingly let AI handle the repetitive parts of development, the industry could shift from manual tuning to a self-improving system.