Unitree Developing AI System for Self-Improving Humanoid Robots

Unitree Robotics is developing an autonomous feedback system intended to let humanoid robots generate code, conduct simulations and improve physical performance independently, founder Wang Xingxing said. Wang introduced the project, Unitree Rehearsal, on Thursday at a physical AI conference that Unitree hosted in Hangzhou. Wang described the project as Unitree's most important undertaking this year. The Hangzhou-based company aims to turn advances in AI models into practical robotics capabilities while reducing manual programming.

Wang said the project would replace labor-intensive programming with a continuous, autonomous development cycle. The proposed system would have AI examine research papers, generate code and test the code in simulations before conducting physical robot trials. Wang said humans could still evaluate performance and supply feedback. Wang said the ultimate aim was an automated cycle capable of accelerating robotics algorithm development exponentially.

Wang described manual coding and debugging by robotics engineers as relatively inefficient. Wang said improving AI models were expected to assume more planning and iteration work, reducing human involvement. Wang said leading AI companies already used models to produce code and collect data with human guidance. Wang argued that robotics companies could adopt a similar approach. However, Wang said robots must also cope with unpredictable environments and mechanical errors, unlike digital software.

Days earlier, Wang predicted that improvements in OpenAI's GPT-6 model could trigger a wider robotics breakthrough within one to two years. Wang's latest speech addressed the steps needed to convert AI model advances into practical capabilities. A video presented at the event showed Unitree robots using GPT-6 for route mapping and movement planning without training for specific tasks. However, the presentation did not establish how reliably the robots could repeat those behaviors in unfamiliar environments.

Wang said large AI models could manage robot planning and navigation. Wang identified the main difficulty as "aligning the AI model's inputs and outputs with those of a real robot". Wang suggested that resolving that problem could allow humanoid robots to tidy rooms on command as early as 2027. Wang offered 2028 as a more realistic estimate.

Tesla CEO Elon Musk told the company's earnings call on Wednesday that humanoid robots could eventually become "the biggest product on Earth" and exceed the human population. However, Musk cautioned that Tesla's Optimus robot still faced roughly a year of development obstacles. Stanford University's 2026 AI Index found that robots completed only 12% of real-world household tasks during evaluations, despite substantially stronger performance in simulation.