Unitree Robotics is developing an artificial intelligence system designed to improve its robots’ capabilities autonomously and plans to incubate more than 200 partner companies over the next five years as it pushes toward practical applications. Founder Wang Xingxing hopes its robots will enter ordinary homes by the end of 2026 or in 2027.
Wang outlined those ambitions at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Summit in Hangzhou on October 22, the company’s first partner conference in its ten-year history. He argued that AI’s potential remains underestimated.
“Almost everyone in the world still doesn’t attach enough importance to AI or have enough faith in it, myself included,” Wang said. Despite the attention AI receives, he believes its pace of development and potential impact remain underappreciated.
Four or five years ago, he recalled, he recommended AI-related stocks to colleagues. “If they had believed more in AI a few years ago, some Unitree employees’ personal wealth might already have increased tenfold or even several dozen times over.”
Wang argued that technical breakthroughs often take five to ten years to reach widespread adoption. He pointed to the roughly decade-long gap between the internet bubble around 2000 and the broader adoption of e-commerce and other applications around 2010. Advanced AI models will also need time to reach ordinary people, he said.
That delay has not diminished his expectations. Wang said he thinks every day about where the biggest opportunities might emerge as AI reshapes people’s lives.
Wang sees the greatest potential in physical AI, systems that enable machines to perceive and act in the real world. As general-purpose AI models improve, he expects a major breakthrough within one to three years.
Testing GPT-6 Astra on robot tasks
Wang said OpenAI’s release of GPT-6 Astra had a significant impact on the robotics industry. Unitree has used the model in several internal experiments.
In one, a robot painted from prompts without the team first writing complex programs for the task, he said. It also changed paints and cleaned its brush on its own.
In another experiment, a robot played Go, recognizing the board and placing stones accurately, Wang said. He repeatedly emphasized that the experiments involved “zero training.”
Robots have traditionally required programming or dedicated training for individual tasks. According to Wang, these experiments suggest that general-purpose AI models could directly control robots for certain activities without additional task-specific training.
He acknowledged the limitations. Models can already handle task planning and navigation reasonably well, he said, but they still make mistakes when manipulating objects.
Picking up an object is relatively straightforward. Assembling precision components is considerably harder, requiring a robot to adjust its movements in real time to account for an object’s position and the forces acting on it. Wang sees substantial gaps in models’ ability to detect changes in their surroundings and correct mistakes promptly.
Even so, he believes advances in large AI models could eventually overcome the main barriers to physical AI.
Wang envisages home robots responding to voice commands to perform tasks such as tidying rooms. His longer-term ambition is for robots to complete around 80% of tasks successfully in 80% of unfamiliar environments, using spoken or written instructions.
Developing a self-improving AI system
Alongside its work with AI model developers in China and overseas, Unitree is developing its own physical AI system. Wang outlined a framework whose Chinese name translates roughly as “Physical AI Robot Self-Evolution V1.0.”
The proposed system would enable AI agents to write code autonomously, run simulations and training, and validate the results on physical robots. Those results would then feed into an automated evaluation process to guide further improvements.
Teaching a robot a new skill has traditionally required substantial work in data collection, training, and optimization. Even after a skill works in the lab, transferring it to another environment may require further adjustments.
Wang wants AI models to automate more of that process, reducing the need for manual programming and debugging while making it easier to develop and deploy new skills.
He expects stronger foundation models to improve robots’ ability to learn and refine their capabilities. Larger training datasets could make that process more effective, while deploying more robots would generate additional real-world data. Over time, he believes this cycle could support an expanding range of skills.
For now, Unitree has primarily disclosed its technical approach. The system’s training efficiency, ability to transfer skills, and success rate in real-world environments remain to be demonstrated.
The project reflects Unitree’s growing investment in the intelligence that controls its machines. Its historical strengths lie in robot hardware, including joint modules, drive systems, and motion-control algorithms. As competition shifts from making robots move to making them perform useful tasks, hardware alone may no longer determine a company’s competitive position.
Wang said Unitree began using pretrained deep reinforcement learning techniques in 2025 and developed an in-house pretrained whole-body motion-control model. The model allows different movements to be trained in batches, substantially improving training efficiency, he said.
“I’m very afraid of wasting a single day”
Wang’s optimism is accompanied by a sense of urgency about competition.
“This industry is extremely active, the future market is enormous, and the prospects are very broad, but the competition is also brutal,” he said.
He described having more things he wants to accomplish each day than time allows, with new technologies to study beyond robotics.
“This is the best of times. I’m very afraid of wasting a single day.”
Wang believes that advances built on thousands of years of human civilization could allow this generation to achieve superintelligence, or AGI, while using humanoid robots to raise productivity.
He also raised a longer-term possibility: as robots take over physically demanding labor and production, changes in how society distributes resources could give more people access to necessities, including housing.
Wang identified four areas critical to that future: large-scale robot manufacturing, the development and deployment of physical AI models, large computing clusters, and the capacity to manufacture computing and memory chips.
Unitree has produced a cumulative 10,000–20,000 G1 robots, he said. Wang considers that scale far from sufficient, arguing that annual production capacity will eventually need to reach tens of millions of robots.
Developing physical AI models and deploying them in practical settings will also be critical to competition. Unitree continues to work with AI model companies in China and overseas while developing its own capabilities.
As models become more capable, demand for computing resources to train and operate robots will increase, Wang argued. Large server clusters, along with the capacity to manufacture computing and memory chips, will be essential to supporting that growth.
In his view, five resources will become increasingly important: data, computing power, chip production capacity, energy, and minerals.
Building a network of partners
Unitree’s dancing, fighting, and running robots have repeatedly drawn public attention in recent years. They have also raised a persistent question: why are the robots performing rather than working?
Wang addressed that question during a roundtable discussion with Yao Ming at the conference. He compared robots boxing or playing ball games with people exercising, arguing that such activities help improve movement, stability, and precision before the machines take on practical tasks.
“If it can’t even walk properly or run, how can you expect a robot like that to do any work?”
Robot fighting, he explained, tests reliability through collisions, falls, and other demanding conditions. It also requires complex movements that place considerable demands on motion-control algorithms and stability.
To move beyond demonstrations, Unitree is working with partners to develop practical applications. Rather than handle all application development itself, the company intends to support other businesses building products and services around its robots.
At the conference, Unitree launched a program to recruit established companies and teams with concrete startup plans. It covers robot hardware and components, embodied AI algorithms, data and simulation platforms, and industry applications.
Wang said Unitree would provide investment, robot platforms, development tools, engineering and manufacturing support, and access to real-world testing and deployment environments. The company aims to incubate more than 200 partner companies over the next five years, expanding its role beyond selling robots to providing a technology platform and developing applications with partners.
The conference was Unitree’s first formal presentation of its broader ecosystem ambitions since going public. With a market capitalization exceeding RMB 100 billion (USD 15 billion), the company is under pressure to translate its founder’s technical ambitions into a sustainable business.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.68 = USD 1 based on estimates as of October 26, 2026, unless otherwise stated. USD conversions are approximate and, where appropriate, rounded for ease of reference. They may not fully match prevailing exchange rates.