Unitree Robotics is developing a “self-evolving” feedback loop designed to allow humanoid robots to write code, run simulations and improve their real-world performance independently, founder Wang Xingxing said, in a move that mirrors a broader push towards “self-improving” artificial intelligence.

Speaking on Thursday at a physical AI conference hosted by Unitree in Hangzhou, Wang unveiled the project, dubbed “Unitree Rehearsal”, which he called the company’s most important initiative of the year.

The project sought to shift robotics development away from labour-intensive manual programming towards a continuous, autonomous feedback loop, he said. Under the system in development, AI would search research papers, write code and test it in simulations before running trials on physical robots.

While humans could still help assess performance and provide feedback, Wang said the ultimate goal was to create an automated cycle that could exponentially accelerate the development of robotics algorithms.

“In the past, every robotics engineer had to write code and debug manually, which is relatively inefficient,” Wang said. “As AI models improve, we expect them to take on more of the planning and iteration, reducing the need for human intervention.”Leading AI companies had already begun using models to generate code and gather data under human guidance, Wang said, arguing that robotics firms could follow a similar blueprint. Unlike digital software, however, robots also had to navigate environmental unpredictability and mechanical errors, he said.

The comments came just days after Wang predicted that advances in OpenAI’s GPT-6 model could spur a broader robotics breakthrough within one to two years. His latest speech focused on the critical steps needed to translate AI model improvements into real-world utility.

In a video shown at the event, robots developed by Unitree used GPT-6 to map routes and plan movements without task-specific training, though the presentation did not demonstrate how consistently the machines could replicate that behaviour in unfamiliar environments.

While large AI models could handle robot planning and navigation, Wang said the biggest challenge was “aligning the AI model’s inputs and outputs with those of a real robot”.

If that challenge could be overcome, humanoid robots could be tidying rooms on command as early as 2027, Wang suggested, though he offered 2028 as a more realistic estimate.

Unitree’s upbeat road map echoed recent comments from Elon Musk. On Wednesday, the Tesla CEO told the US carmaker’s earnings call that humanoid robots could eventually become “the biggest product on Earth” and outnumber humans. He cautioned, however, that Tesla still faced about a year of development hurdles with its Optimus robot.

Data, however, points to a challenging transition from laboratory to living room. Stanford University’s 2026 AI Index found that robots succeeded in only 12 per cent of real-world household tasks during evaluations, despite performing significantly better in simulation.