Unitree's founder puts a number on robotics: two '80%' thresholds
Three weeks after Unitree's Shanghai debut, founder Wang Xingxing used his first public speech since the listing to do something unusual for a robotics CEO on a big stage: he skipped the demo reel and named the metric that decides whether any of this becomes an industry.
Wang Xingxing says the robotics boom starts when a robot handles ~80% of tasks in ~80% of unfamiliar places — and he puts that two to ten years out. Speaking at the 2026 World Robot Conference in Beijing on August 20, days after Unitree listed on the STAR Market, Wang argued the field's real bottleneck is not speed or strength but generalization: models that hit near-100% success in a fixed scene collapse as soon as the object or the room changes. His proposed bar for the field's "GPT moment" is plain — drop a robot into 80% of unfamiliar environments, give it voice or text instructions, and have it complete roughly 80% of the tasks. He said that arrives in two to three years if things go well, or five to ten if they don't.
His diagnosis of why it hasn't arrived is the most useful part. Language models move tokens in and out of a vector space, where a wrong word is just a different word; a robot's every perception, decision, and grasp lands in the physical world with error attached, and the failure concentrates in the last few centimeters. Wang described robots that look like they're about to grasp an object and can't close the final gap — which is why a task that reads as solved in a demo falls apart in deployment. He also pointed at Unitree's own limits: robots can do simple assembly today but slower than a human, and each new task still tends to need fresh training.
The proposed fix is self-evolving physical AI, which Unitree is pre-researching: hand an AI the rules, constraints, and tools, let it search the literature and write robot control code, run that code in simulation, test it on real hardware, then have models and humans score the result and feed it back to the coding agent. It's a closed loop that rides on the improving base models above it — and it's an implicit admission that brute-force human teleop data is not going to scale to general-purpose machines. Worth noting the framing: the man who just rang the bell is telling investors the real inflection is years away, and that the metric to watch is task success in the wild, not a backflip count.
Separately, Chinese AI-infra firm 中科类脑 closed a several-hundred-million-yuan B+ strategic round led by CRRC Capital, the investment arm of state rail giant CRRC. Ginkgo Valley Capital, Shuimu Fund, and Tsinghua-affiliated TusPark funds joined, according to QbitAI and PEdaily. It follows a nine-figure strategic investment in 2025 from China Mobile's Beijing digital-economy fund — so two separate central state-owned industrial funds have now taken positions in the same company within about a year.
The pitch is compute-power coordination: 中科类脑 builds what it calls a "Token factory" with a full energy stack behind it, using power-aware scheduling — it measures value in "intelligence per kilowatt-hour" — to cut the cost per token across mixed domestic and Nvidia hardware. Round proceeds go to three places: inference optimization and unified scheduling across heterogeneous chips, a game-theoretic scheduling brain that balances electrical supply against compute load, and a push to standardize the Token-factory model for replication. Read it as state capital deciding that the binding constraint on Chinese AI is electricity, not silicon — a bet that fits the "East Data, West Compute" grid program, and one where a rail conglomerate is a plausible strategic partner.
Still thin: Unitree shares fell about 15% on the morning after their debut, to roughly 290 billion yuan in market value, after a first-day pop of over 400% that valued a company whose own founder says general-purpose deployment is years away. The listing is the news; what the market does with the timeline Wang laid out is the story to watch.
If a robot has to handle 80% of tasks in 80% of strange rooms before the industry takes off, what does that do to the valuation of companies selling robots today? Tell us in the comments.
