Unitree develops ‘Rehearsal’ to help improve humanoid skills

humanoid.guide reported on October 25, 2026, that Unitree Robotics is developing Unitree Rehearsal, a system to help improve humanoid skills. According to the report, the envisioned process consists of AI-assisted searches for research papers, code generation, simulation testing and trials on physical robots. Bloomberg reported that Unitree already uses feedback from real-world tests to adjust models.

The reported process treats research paper searches, code generation, simulation and real-world testing as stages in a skills-development workflow. For development teams, how each stage is evaluated and how real-world test results are incorporated into model adjustments are specific questions for assessing the system’s effectiveness.

The reported concept centers not on new features for the robots themselves, but on linking the development steps used to create, test and refine skills. The process moves from research paper searches to code generation, simulation and real-world testing. It can be understood as using AI to support the process of taking candidate methods through testing, rather than relying solely on people to tune individual movements. However, the reports do not say which stages will be automated or who will validate the generated code. The “model adjustments based on feedback from real-world tests” described by Bloomberg should also be distinguished from the Rehearsal development concept. The reports do not say how much the method said to be already in use overlaps with the system under development. If a process is established for feeding real-world results back into models, it will be important to measure how simulation assessments differ from actual performance and how to use those findings in subsequent tests. But the reports do not specify the types of training data, evaluation criteria, scale of real-world testing or safety review procedures. We have covered individual Unitree robots, including humanoid.guide’s introduction to Unitree Robotics’ G1+ and humanoid.guide’s introduction to the “G1-D”. While those stories focus on robots, reports on Rehearsal shift the focus to the skills-development process. However, based on the article summaries alone, there is no confirmation that Rehearsal will be used with specific robots such as the G1+ or G1-D, so a connection between them cannot be assumed. The industry implication is how to establish the cycle between simulation and physical robots as a development process. Success in simulation does not guarantee reproducibility on a physical robot, and feeding real-world test results back into a model alone does not establish that a skill is stable or safe. The next points to watch are Rehearsal’s implementation scope, its evaluation methods and the specific mechanism for feeding real-world test results back into model adjustments.