❯ Unitree develops a robot self-improvement system as G1 production reaches an estimated 10,000–20,000 units

Closing the training loopAt the October 22 World Physical AI Conference, Wang Xingxing said Unitree is developing “宇树预演,” a robot self-improvement system. Conference coverage describes AI collecting papers, writing code, training in simulation and testing on real robots, with scores fed back into the next iteration. The system remains under development.

Production and capabilityWang said roughly 10,000–20,000 G1 humanoids have come off production lines since the model’s 2024 introduction. Production is not a count of delivered robots or commercial deployments. He also showed GPT-6-based robot task tests; The Paper’s on-site report records painting and navigation examples.

From instructions to movementWang expects more substantial embodied-AI progress over the next one to three years and wants robots to perform a larger share of everyday tasks from language instructions. Those are forecasts and goals. The engineering challenge is converting model understanding and plans into precise, reliable interactions with varied physical objects.

Real-world feedback as training dataA closed loop tests simulated solutions on hardware and feeds successes and failures back into training. More robots provide more potential platforms, but useful data still requires well-defined tasks, sensing and action feedback. Connecting models to physical performance is central to the next development stage.

▮ SIGNALProduction supplies the hardware base; further capability gains require turning physical successes and failures into training feedback.