Unitree is paying to pull outside developers onto its hardware, betting that a wide pool of builders can close its gap in robot AI models faster than in-house work alone.
Unitree Robotics will hold a global competition for robot AI every year and fund it with tens of millions of yuan annually, founder and chairman Wang Xingxing said on October 22. He announced the World Physical AI Competition (世界物理AI大赛) at the company’s 2026 World Physical AI Conference and partner ecosystem conference in Hangzhou, IT Home reported, citing Securities Times.
Unitree will hold a World Physical AI Competition every year, with prize money of tens of millions of yuan and eight tracks open to teams and individuals worldwide. The tracks cover free movement, track and field, fighting, dexterous manipulation, production work, sports events, multimodal interaction and elderly and health care. IT Home did not report a schedule, rules or judging criteria.
Wang used the same speech to set out where Unitree will put its money next: building production capacity for very large volumes of robots, developing small physical-AI models suited to its own hardware, and building very large computing clusters. The plan points to the gap Unitree has acknowledged in robot "brains", as opposed to bodies, an area where it has lagged its own reputation in hardware and motion control, according to Yicai.
Three priorities aim at volume, models and compute
According to IT Home’s account, Wang said Unitree must scale manufacturing to meet future demand that could reach tens of millions to hundreds of millions of robots. On models, the company will both build its own and work with partners on small physical-AI models matched to its machines. Yicai reported that Unitree is cooperating with large-model companies in China and abroad. Wang’s list in Yicai’s version also included manufacturing chips for computing and storage.
Unitree is also building what Wang called a robot self-evolution system. AI models gather papers, write code, train robots in simulation and then test them in the real world, with AI, humans or both scoring the results to drive the next round of changes, Yicai reported. The company says the loop is meant to replace manual tuning and speed up algorithm development, and Wang told the 21st Century Business Herald that it is gradually taking shape.
Wang puts a one-to-three-year clock on a breakthrough
Wang’s talk, about 30 minutes long, carried the title "Believe, Believe in AI as Much as Possible", Yicai reported, the second time in ten days he has made the point publicly. He repeated that his biggest mistake of the past decade or more was not believing enough in AI, a theme EastFrontier covered after his October 13 remarks on GPT-6.
He told the Business Herald that he could already see signs that physical AI was about to break through, and that it could happen within one to two years, or two to three. In his view, robot motion planning and navigation are largely solved. The hard part is aligning what models take in and put out with the physical world, which is why success rates fall when a robot tries to grasp an unfamiliar object.
Wang described the target in concrete terms, according to Yicai: a robot that, given a voice or text instruction, can complete about 80% of tasks in about 80% of unfamiliar settings. He also described an internal test in which GPT-6, with no task-specific training, controlled a robot to paint, switching colors and washing its brush in a nearby tray without being programmed to do so.
The contest is a bet on Unitree’s developer base
Wang credited much of the industry’s progress since the second half of last year to developers building algorithms and open-source frameworks on Unitree machines, the Business Herald reported. A global contest with large prizes is a way to direct that outside work at the tasks Unitree most needs solved, such as dexterous manipulation and factory production, on Unitree’s own hardware.
The competition’s rules, schedule and first tracks to open will show how much of that work Unitree can capture, and whether entrants must use its robots. Unitree’s third-quarter report will show how much it is already spending on models and compute.