"It's not that it can't go fast. Go faster and the error rate rises, and it drops the ball during demos."
That is Pan Jia, an iFlytek distinguished scientist and a lead contributor on a national science-and-technology prize, explaining why the robots at this year's World Artificial Intelligence Conference in Shanghai moved so sluggishly. The slowness wasn't a hardware limit. It was a safety gear, deliberately engaged.
Everyone who walked the WAIC floor asked the same question—it's 2026, why are the robots still so slow?—and reached for the same comforting answer: the tech just isn't ready yet. Pan's admission flips that read. The sluggishness is a confession. The body is fine; the supply chain for joints, motors and dexterous hands gets more mature every year. The brain hasn't crossed the usability threshold, so the whole floor is performing a maturity it doesn't have.
The body is fine; the supply chain matures every year. The brain hasn't crossed the usability threshold, so the whole floor is performing a maturity it doesn't have.
Change the environment or the object, and the motion starts to deform. That's why so many showcases stay pinned to the demo stage, executed carefully inside a scene arranged just so.
This is the fact the "ChatGPT moment" narrative keeps papering over. There is no converged race here to be behind on—and the clearest proof came from a stage where six labs couldn't agree on almost anything.
At WAIC's "World Model Six Dragons" forum on July 19, six technical leads sat down and openly split on every load-bearing question in the field: what a world model is, which architecture wins, how to train it, what data standard to use, how to even measure success.
Route down the middle: pixel-generation (Genie Vision), latent-space JEPA (Wujie Power), and a fusion camp that includes AgiBot, Ant Lingbo and others. That's three answers to the single question of what space a model should reason about the world in.
Then came the sharper clash. How faithfully must a world model understand physics?
Wujie Power's Xia Zhongpu wants prediction precision on geometry, motion and force—a model graded like a physics simulator. Ant Lingbo's Shen Yujun countered that "physical plausibility" beats "physical precision": a robot needs to know that iron falls faster than cotton, but not by exactly how much. It just needs to observe.
智元's Ren Guanghui called long-horizon physical consistency an indispensable metric. 自变量's Wang Hao called long-horizon reasoning a "pseudo-need"—if inference is so slow it misses the decision window, no amount of understanding matters.
Notice what each argument is actually defending: the home team's strength. Latent space can regress physical quantities directly, so Wujie dares demand precision. Pixel generation naturally learns "looks plausible," so Genie Vision promises only plausibility. 智元 sells its world model as a simulator, and long-horizon rollout is a simulator's lifeblood; 自变量 fuses prediction into an end-to-end action pipeline, where every extra step of rollout eats decision time.
These aren't disagreements about truth. They are position papers for capabilities each lab already owns. And that is exactly what a pre-paradigmatic field looks like.
The tempting response is that China is simply trailing the West, still arguing while American labs ship the real thing. The evidence says otherwise.
At RSS 2026 in Sydney, a source from Sergey Levine's lab told Qbit AI something that should recalibrate the entire conversation: the leading North American embodied-AI companies, Physical Intelligence included, "are still a fair distance from maturity, and the field has yet to produce truly foundational work."
The same source added that China's real lead is in hardware—so why, he wondered, do Chinese companies keep branding themselves the "Chinese version of X"?
A Physical Intelligence intern was even blunter. On the popular claim that a world model can predict the consequences of an action and steer the robot away from mistakes, he asked: "Have you actually seen it do that? Have you actually seen a robot with a deployed world model say, this action might be a problem, I shouldn't do this, I should do something else?"
The Qbit reporter had no answer. Neither, really, does the field.
So if the architecture wars are a coin flip and even the acknowledged leaders admit no foundational work exists, where is consensus actually forming? Not on a theory. On a benchmark.
Tao Dacheng, moderating the Six Dragons forum, offered the sharpest framing of the day. The deep-learning explosion, he noted, didn't happen because neural networks won a debate against feature engineering. It happened because ImageNet dragged every method onto the same ruler. Standardize the measure, and convergence follows.
Debate is fodder for papers, Tao said; overlap is the bedrock of industry. His lab used the forum to release Physical IQ, pitched as the first embodied-native benchmark for physical intelligence across bodies, tasks and scenes. A shared ruler, not a winning theory.
The second quiet consensus is even less glamorous than a benchmark: data. Not more of it—better curation of it.
At RSS, EgoVerse's Simar Kareer, who has interned at Physical Intelligence, argued that what separates models isn't raw volume but how data is filtered, mixed and organized—the same lesson OpenAI and Anthropic learned in language. Asked how to scale embodied data, he advised against simply buying hours from data vendors.
His own experiments make the point concrete: robot data plus free-form human video barely helps; task-aligned human demonstrations alone barely help. The jump comes only when all three appear together, giving the model enough "aligned regions" in the data mixture. Diversity isn't pouring different sources into one pool.
The architecture-agnostic bet keeps recurring. iFlytek's new spin-off, Ta Fang Intelligence—Pan Jia's outfit, capitalized at 300 million yuan—chose video generation for its predictor not because it's the most advanced route but because it fits the backbone. Pragmatism over allegiance.
And Shen Yujun's own convergence bet isn't an architecture at all. It's touch: the moment tactile data breaks through, he argues, world models for the physical and digital worlds part ways for good.
That's where the unglamorous moats are being dug. NeoteAI, with Fudan, just released NeoData—over 30,000 hours of vision-tactile interaction, 1.4 million clips, across six robot bodies and 450 real long-horizon tasks. Their pitch: touch is the next scaling law, and their tactile model lifts a plug-insertion success rate from 60% to 85%, key-pulling from 35% to 99%.
None of this is a language-model-style ChatGPT moment. It's plumbing.
Which is precisely why it matters for anyone writing checks. Caixin counts roughly 370 startups that entered China's embodied-AI field over two years, with about 50 chasing Hong Kong or mainland listings by July.
That capital is racing a pre-paradigmatic field. Betting on an architecture today—pixel versus JEPA versus fusion versus Pan's "body cognition"—is betting on a coin flip, because the people closest to the frontier admit no one has proven which flips heads.
Wang Hao offered the number that should haunt every deployment deck: pushing a model from 90% to 99%, then 99% to 99.9%, does not cost linearly. A tiny error rate in the lab becomes, on-site, a stream of human takeovers and reworks—a bill that stops adding up.
Or as the forum distilled it: a demo shows the capability ceiling; deployment tests the capability floor. The money is chasing demos of an upper bound. Deployment lives or dies on the lower one.
The safest read, then, isn't which lab has the right architecture. It's which one is building the boring layers—the 30,000 tactile hours, the alignment findings, the evaluation standards—that will still matter after the architecture debate collapses onto somebody's ImageNet.
Because when the ChatGPT moment finally arrives, it won't settle the argument. A shared ruler will. And the labs quietly building that ruler, and the data to feed it, are the ones worth watching—not the ones with the fastest demo they're too nervous to run at full speed.