Wang Xingxing says to believe in AI; Unitree's stock price rebounds over 54%, but its self-developed model still awaits on-site proof

In recent days, Unitree Technology has reversed its previous slump and risen continuously. The latest stock price has rebounded by more than 54% cumulatively from the intraday low of 404.3 yuan per share on October 15. In particular, today it hit the 20-centimeter daily limit and closed at 624 yuan per share, with its market value once again exceeding 252.4 billion.

Image | Source: Wind

Along with the surge, Unitree has just released the H2-X Kung Fu Legend robot positioned as a "bodyguard." In the demo video, it lifted a car and performed flying kicks and elbow strikes. In terms of load capacity, it has made great progress compared with its previous dancing form, and it is clearly still trying to break through the limits of physical hardware.

And at the just-concluded 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Wang Xingxing devoted almost more than 90% of his speech to talking at length about AI.

He said, "The whole world, including myself, still does not have enough faith in AI"; "If colleagues paid more attention to AI, their net worth might grow tenfold or even dozens of times"; "There may be only a few thousand people who truly understand the potential of cutting-edge AI models"; "Physical AI has already shown signs of being broken through, in as fast as 1 to 2 years, or as slow as 2 to 3 years..."

This is almost completely different from Wang Xingxing a few years ago. Since founding the company in 2016, Wang Xingxing has spent more energy and thought on studying hardware.

Judging from the demonstrations at this conference, Unitree is already advancing self-developed physical AI models and a robot self-evolution system, but the most intuitive on-site demonstrations, such as drawing and playing Go, are mainly based on test results from GPT-6 Astra.

The outside world may be more concerned about when Wang Xingxing, who has become an AI believer, will actually deliver Unitree's own AI capabilities. As the hardware capabilities of humanoid robots continue to improve, whether Unitree can turn self-developed AI from a technical route into real product capability will become an important test in the next stage.

In the past two days, Unitree held its first ecosystem conference in ten years in Hangzhou. The newly released Unitree H2-X is very eye-catching: it can easily flip over a car, its flying kicks and elbow strikes can dent car doors, it can lift a car with its bare hands, and it can jump from a 4-meter-high platform and land steadily. The company claims a maximum battery life of 8 hours.

Image | Source: Unitree official video account

The theme of the main forum at this conference was "World Physical AI Conference," and the title of Wang Xingxing's speech was "Believe, Believe in AI as Much as Possible." He devoted almost the entire speech to AI: the world still does not have enough faith in AI, including himself; four or five years ago he reminded colleagues to pay attention to the stocks of leading AI companies, and if they had believed at the time, their personal wealth might have grown tenfold or even dozens of times; there may be only a few thousand people who truly understand the potential of cutting-edge AI models; physical AI has already shown signs of being broken through, in as fast as 1 to 2 years, or as slow as 2 to 3 years...

Image | Wang Xingxing's speech

First of all, it must be said that Unitree does have self-developed AI products. In September, Unitree released the general humanoid robot foundation model UnifoLM-WLA-1.0. At this conference, Wang Xingxing further introduced the "Physical AI Robot Self-Evolution V1.0" that is under preliminary research.

He also made "Physical AI Robot Self-Evolution V1.0" the most important preliminary research direction of this year—allowing AI to independently search papers, write code, train in simulation environments, then enter real robot testing, and form a continuous iterative loop through AI or human scoring feedback.

But a preliminary research framework is not the same as a model that has already been made to work, and a technical roadmap is not the same as capabilities that have already been implemented.

The most intuitive AI demonstration on site actually came from GPT-6 Astra: under zero-training conditions, the robot completed drawing, washing brushes, changing paint, playing Go, and indoor navigation. Wang Xingxing himself admitted that large models can already solve planning and navigation relatively well, but they are still prone to errors when actually performing manual operations.

This shows that external large models can indeed drive robots to complete some tasks, but it also exposes that Unitree's self-developed model has not yet come to the fore.

More importantly, UnifoLM-WLA-1.0 was released more than a month ago, but at this conference there was not a single on-site hands-on demonstration based on that model. What the audience saw was still footage driven by GPT-6 Astra, while Unitree's own model still lacks real, practical implementation.

It has already built the skeleton—models, data, evaluation, and the open-source ecosystem are all being advanced—but it is still some distance from "letting a robot understand one sentence and then go do the work." Wang Xingxing predicts that the "ChatGPT moment" for embodied intelligence will arrive as early as 2027, which means Unitree has left itself a window of more than a year to catch up.

The problem is that during the window period, competitors will not stop and wait.

Hardware can continue to iterate, but intelligence is the next hurdle

Unitree's hardware advantages are real. The H2-X can lift a car, perform flying kicks, jump down from a 4-meter-high platform, and has 8 hours of battery life. Its motion control capabilities are in the first tier among humanoid robots globally.

But hardware has no moat in China. Motors, reducers, controllers, and LiDAR—the supply chain is already mature, and latecomers can quickly catch up through engineering iteration. If you happen to have visited the 2026 World Robot Conference, many people's first impression was: humanoid robots that can sing and dance are everywhere, and this is not the unique skill of any one company.

Image | Source: World Robot Conference; the scene looked like a vegetable market

Therefore, relying on continuous hardware improvement to maintain a certain level of advancement is far from enough for Unitree.

What truly determines who can stay at the table is intelligence.

Wang Xingxing himself put this hurdle very clearly. He set a quantitative standard for the "ChatGPT moment" of embodied intelligence: in about 80% of unfamiliar scenarios, complete about 80% of major tasks through natural language instructions. This standard is not low—it means robots must go from being "demo products in specific scenarios" to "usable products in general environments."

The difficulty is that large models make robots better at "thinking," but doing housework is not chatting.

Wang Xingxing admitted that large models can already solve planning and navigation fairly well, but they are still prone to errors when it comes to actual manipulation. The core bottleneck is the alignment between the input and output of AI models and real robots: sensor readings have deviations, joint responses have delays, and grasping force, friction, occlusion, and lighting changes all introduce errors.

Picking up a chess piece may not be difficult, but to assemble small parts into place and safely put fragile cups into a dishwasher, the final few millimeters of error can determine success or failure.

What is more troublesome is generalization. Current AI models, after sufficient training in fixed scenarios, can approach a 100% success rate; but if the position or shape of an object or the environment changes slightly, the success rate may drop significantly.

Home scenarios are precisely the least "cooperative": desktops are different every day, elderly people, children, and pets change the state of the space at any time, and tasks are often vague—"tidy up the living room" sounds simple, but behind it are recognition, planning, grasping, moving, obstacle avoidance, error correction, and failure retries. A robot cannot just complete one beautiful demonstration; it must succeed repeatedly in unfamiliar environments.

This is exactly what Unitree's self-developed model needs to prove.

UnifoLM-WLA-1.0, released in September, is a 6B-parameter embodied multimodal large model, trained on about 2,500 hours of real-machine data, with a single model coordinating 64 tasks and leading open-source models in multiple embodied reasoning benchmarks.

Image | Wang Xingxing explaining the operating mechanism of Unitree's self-evolution system

But leading in benchmarks does not mean it can perform on site. UnifoLM-WLA-1.0 has been released for more than a month, yet at this conference there was not a single on-site hands-on demonstration based on the model. What the audience saw was still footage driven by GPT-6 Astra.

This precisely shows that Unitree's self-developed model still lacks an "on-site proof." Releasing a model, proposing a technical route, and building a computing cluster are one thing; proving the model's actual capabilities is another. What is truly worth watching is whether the self-developed model can stably complete tasks on real robots, whether it can adapt to unfamiliar environments, and whether it can continuously improve based on execution results.

The prospectus shows that Unitree's total R&D expenses for the three years from 2023 to 2025 were about 265 million yuan. Compared with the R&D investment of AI large-model companies, such funding is not on the same order of magnitude at all. Wang Xiaochuan once revealed that the single training cost of a large model with 100 billion parameters is between 30 million and 50 million yuan. And several practitioners in the embodied intelligence track also lamented to Phoenix Technology that investment in embodied brains is enormous, with daily consumption at the million-yuan level.

By then, when people evaluate a robot, they will probably no longer ask whether it can carry a car, but will ask: can it clean up the kitchen before I leave the house, and do so without scaring the cat away?

This article is from the WeChat public account "Phoenix Technology," author: Lu Chunfeng, published with authorization from 36Kr.