Unitree Robotics is advancing a “self-evolving” feedback loop designed to enable humanoid robots to autonomously write code, run simulations, and enhance their real-world performance, according to the company’s founder, Wang Xingxing. The development marks a significant shift in robotics toward greater automation in the programming and iterative improvement process.
Wang introduced the “Unitree Rehearsal” project at a physical artificial intelligence conference held by Unitree in Hangzhou on Wednesday, describing it as the company’s most important initiative this year. The system aims to move robotics development away from the traditional, labor-intensive manual coding methods toward an autonomous, continuous feedback mechanism.
Under the new approach, artificial intelligence would autonomously research relevant papers, generate code, and test it through simulations before deploying it on physical robots. Human engineers would still contribute by evaluating performance and providing feedback, but the ultimate goal is to establish an iterative cycle largely independent of direct human intervention. Wang noted that this would significantly accelerate the evolution of robotics algorithms.
“In the past, every robotics engineer had to write code and debug manually, which is relatively inefficient,” Wang explained. “As AI models improve, we expect them to take on more of the planning and iteration, reducing the need for human intervention.”
Wang also highlighted that while leading AI firms are already using models to generate code and collect data under human oversight, robotics faces additional complexities. Robots must contend with mechanical failures and environmental unpredictability, challenges not present in purely digital software development.
The presentation included a video demonstration showing Unitree robots utilizing capabilities from OpenAI’s GPT-6 model to map paths and plan movements without task-specific training. However, Wang did not provide information on the consistency of this performance across unfamiliar environments. He acknowledged that the key hurdle remains aligning the inputs and outputs of AI models with the operational demands of physical robots.
If this challenge is overcome, Wang projected humanoid robots capable of performing tasks such as tidying rooms could be commercially available by 2027, with 2028 deemed a more cautious estimate.
Unitree’s vision aligns with recent statements from Tesla CEO Elon Musk, who on the company’s earnings call noted that humanoid robots might eventually become the largest product line globally and potentially outnumber humans. However, Musk also warned that Tesla’s own humanoid robotics project, Optimus, still faces approximately a year of development challenges.