Unitree puts G1 output at 10,000 to 20,000 units and wants its robots to test AI-written code

Unitree founder Wang Xingxing told the company's 2026 World Physical AI Conference in Hangzhou on Oct. 22 that cumulative production of the G1 humanoid robot, released in May 2024, has reached about 10,000 to 20,000 units. On the same stage he described a project he called one of the most important things Unitree did this year: pre-research on "Physical AI Robot Self-Evolution V1.0," which overseas reports have called Unitree Rehearsal.

Unitree is betting that its edge in robot intelligence will come from the number of robots it has sold and can still build, and that it does not need to own the best model. The loop borrows outside frontier models for the thinking and needs real robots for the testing, which is where a G1 fleet of that size comes in.

The model is borrowed and the test bench is Unitree's own

As Wang described it, Unitree defines the skills, tools and task constraints. A frontier AI model then searches recent papers and open-source code and writes the robot's control code, which is trained and checked in simulation before it runs on physical robots. The AI model and human reviewers score the result, and the score goes back to the coding agent for the next round.

Wang listed four conditions for the loop: a stronger foundation model, multi-source data at scale, more real-robot deployments, and skills that accumulate into a scaled skill ecosystem. Three of the four are data, robots in the field and accumulated skills.

He also gave an example of what a borrowed model can already do. GPT-6 has been deployed on Unitree robots with zero training, he said, and can paint and wash its brushes, play Go and Chinese chess on a fine board, and find a kitchen and look for tools. "We did not write programs specifically for these actions," he said, according to RobotJia.

The three directions Unitree set out at the event fit the same bet: mass-production capacity for future demand in the tens or hundreds of millions of units, small physical AI models built in-house and with partners for its own hardware, and very large computing clusters.

Wang says the gap between model and robot is still open

The strongest case against the bet comes from Wang himself. The loop is labelled pre-research, and he said planning is largely solved while the biggest unsolved problem is alignment error between an AI model's input and output; contact and interaction with real objects is not yet handled well. If that error does not shrink, a larger fleet mostly adds more failed trials to score.

He defined embodied AI's "ChatGPT moment" as a robot completing about 80% of tasks from voice or text instructions in 80% of unfamiliar scenes, and put it next year at the earliest and about five years away at the latest.

For scale, China's humanoid robot shipments passed 40,000 units in the first half of 2026, according to Ministry of Industry and Information Technology data cited by Zhidx. Unitree's prospectus put its own 2025 humanoid shipments above 5,500.