Unitree wants to use AI to help its robots learn to improve themselves

Unitree Robotics presented, at a physical AI conference in Hangzhou, an “self-evolving” development system that uses AI agents to search research, write control code, test it in simulations and select results for validation on real robots. The proposal is to automate part of the cycle currently handled by engineers and allow robots to acquire new skills faster and at greater scale. According to Xataka, the project is called Unitree Rehearsal.

Wang Xingxing, the company’s founder, told the South China Morning Post that more advanced AI models could take over a larger share of planning and iteration. Unitree has also been testing OpenAI’s GPT-6; in internal experiments, Wang said, robots performed tasks such as painting and playing Go without specific training. Xataka notes, however, that the method’s efficiency, skill transfer and consistency in unfamiliar environments have yet to be demonstrated. The models handle planning and navigation better, but still make mistakes when manipulating objects.