Unitree is developing a "self-evolving" AI system so its humanoids improve themselves
From the Robots digest for 2026-10-23
What happened: Unitree Robotics, the Hangzhou-based humanoid maker, is developing a "self-evolving" feedback AI system it says will reduce the need for manual programming and turn advances in AI models into real-world robot capability. The report was published around 12:45 AM US Eastern on October 23, 2026. A midday pass (~12:48 PM ET) adds a founder-attached performance target: Unitree's Wang Xingxing now expects humanoid robots to complete 80% of spoken tasks by 2028, with the self-improving AI named as the route there.
What's genuinely new: Unitree's first named program at the software layer rather than the hardware layer: instead of another robot body — the H2-X "guardian" humanoid shown in a promotional video on October 22 disarming attackers, lifting cars, and rescuing people — this is a claimed system by which the robots improve themselves, cutting manual programming out of the loop. The program now carries its first measurable definition — 80% of spoken tasks completed by 2028 — but it still arrives without a system name, a mechanism description, or a demonstration; an announced program with a dated ambition, not shipped technology.
Watch: the feedback mechanism's specifics, an independent demonstration that a Unitree robot improves without re-programming, measured progress against the 80%-of-spoken-tasks 2028 target, and which of the company's robots — the ~$16,000 G1 or the H2 line — actually consumes the system in a shipping product.
Links: MSN · The Next Web