Unitree Robotics is building a system called Unitree Rehearsal, meant to let its humanoid robots improve themselves, and its founder ranks it above everything else the company is working on this year.
Founder Wang Xingxing described the project at a physical AI conference that Unitree hosted, according to a South China Morning Post report by Minxiao Chang. He called Rehearsal the company's most important project of the year. That is what the account we have confirms: a name, a goal, and a priority ranking from the person who sets the priorities.
"Physical AI" is the industry's term for AI that controls machines acting in the real world, as opposed to chatbots producing text. Self-improvement is a large promise in that field. Humanoid robots are often taught by people, through hand-written control software, recorded demonstrations, or training in simulation, followed by software updates. A robot that gets better without that loop would lower the cost of each new skill.
The phrase "improve themselves" can cover very different things, though. It could mean practicing in simulation, learning from the real-world experience of many robots, or something looser. The name Rehearsal suggests practice before performance, but that is a reading of a label, not a described method.
The ambition is not unique to Unitree. Rivals such as Tesla, with its Optimus robot, also pitch learning as the route to useful humanoids. What sets this announcement apart is the framing: the founder of a robot maker put self-improvement at the top of the company's list. The useful test is whether Unitree shows a specific task where a robot measurably gets better without engineers retraining it by hand.