--- title: "Unitree's Founder Drew a Line: 80% of Places, 80% of Tasks" date: 2026-09-10 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-wang-xingxing-80x80 language: en --- # Unitree's Founder Drew a Line: 80% of Places, 80% of Tasks > Wang Xingxing says embodied AI's real bottleneck is generalisation, and proposes a concrete threshold — a robot that works in 80% of unfamiliar places on 80% of tasks. That is 64%, and the industry is not there yet. A four-year-old visiting someone else's home for the first time can find the bathroom, turn on the tap, and drag a stool over to reach the sink. Nobody taught that child this particular floor plan. The most advanced robots in the world cannot do it. On **9 September 2026**, Wang Xingxing — the founder of Unitree Robotics — put a number on that gap, speaking at a conference on private-sector innovation in China. ## Key takeaways - **The bottleneck is generalisation, not hardware.** Wang's assessment: today's embodied AI fails when the environment changes, not when the task is hard. - **A quantified threshold:** a robot counts as useful when it can complete **80% of tasks in 80% of unfamiliar environments** — an effective success rate of **64%**. - **Ambition, not promise:** Wang said he hopes robots reach self-improvement within one to two years. He framed it as a hope, not a commitment. - **Industry timeline from a peer founder:** Huang Guan (Jiaji Vision), speaking at JD's JDD conference the same week, said leading companies broadly agree on **2–3 years** to a robotics "ChatGPT moment" for ordinary tasks, **around 5 years** for complex long-horizon tasks, and **5–10 years** before robots enter ordinary homes. ## Why "unfamiliar" is the hard part Robots are already good at fixed actions in known environments. Put a machine at one workstation, with one part and one motion, and repeat it a few thousand times: reliability gets very high. Unfamiliar is different. It means the lighting is not what the camera was trained on. The floor has different friction. The door handle is round instead of lever-shaped. The mug you want is behind a laptop. Humans handle all of this with common sense, which is really just several decades of compressed experience. A robot is not short of compute; it is short of those decades. Wang's prescription follows directly from the diagnosis: **deploy more robots in the real world.** Not more laboratory experiments — more hours in actual places, accumulating the data that only reality generates. His analogy is learning to drive: ten thousand hours in a simulator is worth less than a hundred kilometres on real roads. ## The sequencing matters: factories first, homes last The 5–10 year estimate for domestic robots is the number worth sitting with, because the ordering is not accidental. Factories are standardised: the same position, the same motion, the same acceptance criteria. Homes are not. "Clean" means one thing to you and something else to your mother — and no training set resolves that disagreement. This is also why the current commercial reality is lopsided. Unitree's own listing documents showed that in the first nine months of 2025, **more than 70% of its humanoid revenue came from research and education**, with genuine industrial application at roughly **9%**. ## What this means if you are not building robots - **Postpone the home-robot expectation.** Products marketed as "home robots" today almost all do one thing — vacuum, mow, hold a conversation. Folding laundry and cooking remain years away on the industry's own timeline. - **If you work in a factory, warehouse or industrial park, the window is open now.** Standardised environments are what gets automated first. Learning to work alongside these machines — or maintain them — is the highest-return response. - **Stop using backflips as the progress metric.** A backflip demonstrates control. Generalisation determines usefulness. The right question when watching a robot video is simple: would it still work in a different room? - **The jobs in the gap are not research jobs.** Between "built" and "useful" sits installation, commissioning, scenario adaptation, maintenance and data collection — roles that need industry knowledge more than a PhD. Whether a robot can do your chores does not depend on whether it can do a backflip. It depends on whether it can do an unpractised task in an unvisited place. *Remarks as reported from the 9 September 2026 conference and JD JDD 2026.* --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-wang-xingxing-80x80 Free to quote with attribution and a link to the original.