Unitree is building self-improving humanoid robots

Unitree is developing an AI system intended to let its humanoid robots improve their own software and simulations, turning robot development into a more automated feedback loop. The project, called Unitree Rehearsal, was outlined by founder Wang Xingxing at an AI conference hosted by the company, according to reporting by TechRadar.¹

The concept is more specific than the usual promise that robots will become “smarter”. Unitree says the system could allow robots to write code, run simulated tasks and use the results to improve the algorithms controlling their physical behaviour. Human engineers would still provide feedback and assess performance, but the company’s stated aim is to reduce how much manual planning and debugging is needed for each iteration.¹

That distinction matters. A robot learning from demonstrations is not the same as a robot independently rewriting the software that governs its movement. Unitree’s proposal sits somewhere between the two: people remain responsible for judging whether a result is useful or safe, while the system handles more of the repetitive work involved in testing and refining a controller.

What Unitree is proposing

Wang described the existing robotics workflow as inefficient because engineers must write code, run tests and debug failures by hand. Unitree Rehearsal is designed to automate more of that cycle. In practice, the system would use simulations to rehearse behaviours before updated software is deployed to a real humanoid robot.¹

The company has linked the idea to practical domestic tasks. Wang said Unitree robots could be cleaning rooms as early as next year, although that is a prediction rather than a confirmed product launch date. The gap between a robot completing a carefully staged demonstration and reliably handling an untidy home remains enormous: floors, furniture, lighting and objects change constantly, and a mistake can damage property or hurt someone.

Unitree’s official G1 material shows the sort of platform the company is developing around AI-assisted movement. The G1 is presented as a humanoid robot with imitation and reinforcement learning features, force-controlled hands and a robot world model, although that product page does not itself confirm that the new Rehearsal system is available on the G1.²

The wider humanoid race

Unitree is not working in isolation. TechRadar’s report also points to fresh developments from Tesla and Figure.¹ Tesla says its Optimus production lines are operating in Fremont, with Elon Musk projecting a possible ramp towards 20,000 robots per week by 2027. That target still depends on further engineering work, and Tesla acknowledges that Optimus cannot yet complete varied tasks independently without careful coordination between perception, planning and movement.

Figure, meanwhile, says its roadmap has changed substantially over the past year. The company is working towards a fourth-generation humanoid and claims that its Figure 03 production run has reached four figures. It has also highlighted millimetre-level fingertip tracking and a large video dataset gathered through its Index app.¹

Those announcements point to three different bottlenecks. Unitree is focusing on automating the software-development loop; Tesla is talking up manufacturing scale; and Figure is emphasising dexterity and training data. None of those advances, on its own, proves that general-purpose humanoids are ready for ordinary homes.

Our opinion

Unitree Rehearsal is interesting because it targets the least glamorous part of robotics: the slow, expensive repetition required to make a machine behave consistently. If simulations can genuinely shorten the write-test-debug cycle, the approach could help companies improve robots faster without putting every unfinished change straight into a physical machine.

But “self-evolving” is an easy phrase to oversell. A system that proposes code or simulation changes is still operating inside goals, data and safety limits chosen by people. The crucial test will be whether Unitree can show repeatable improvements on unfamiliar tasks, with failures caught before they reach a person or a real building. Until that evidence exists, Rehearsal should be treated as a promising development programme rather than the beginning of an autonomous robot revolution.

For now, the news is less about robots secretly taking over and more about manufacturers trying to industrialise the painstaking process of teaching them. That may be less cinematic than The Matrix, but it is the part most likely to shape whether humanoids become useful products or remain impressive demonstrations.