Two Robot-Brain Labs Just Made Opposite Bets on Whether a Robot Should Know Its Own Body

Demos have gotten smooth. Deployments still look broken. The robots that dazzle on stage can also freeze when a sleeve catches a corner they did not see coming.

The gap, according to a new Chinese lab spun out of iFlytek (a voice-AI company headquartered in Hefei), is that today's robot brains are not being trained to know their own bodies. They are only being trained to react.

The lab is 爻方 (Yaofang) Intelligence, registered in late June with 300 million yuan in registered capital (roughly $42 million at recent exchange rates, approximate). It is led by Pan Jia. The company made its public debut at WAIC 2026, the World Artificial Intelligence Conference in Shanghai, and used the moment to put a stake in the ground on a question the field has been ducking.

The dominant recipe in robot brains is called VLA, for vision-language-action. It is the same architectural idea as the large language models behind ChatGPT, but the output is motor commands instead of text. A camera frame goes in, a sentence like "pick up the mug" goes in, and the model emits joint torques out. VLA is good at reflexes. It is bad at consequences. When the gripper slips, the model has no internal picture of what just happened, so it has nothing to roll back to.

The next layer the field has been reaching for is the world model: a system that imagines the next few seconds of physics before acting. World models are how robots could plan, not just react. But they have two exposed failure modes, and 爻方's technical report, posted to arXiv as iFLYTEK-Embodied-Omni (2607.02542), names both. First, error compounds: small prediction mistakes pile up over a long plan, and the imagined future drifts away from the real one. Second, and this is the one Pan Jia is hammering, the model has no proprioception, the sense of where the robot's own joints, links, and gripper are in space.

The doorway matters here. Knowing where your hand is when your eyes are closed is proprioception. For a robot, that means knowing whether its arm can reach the mug at all, how much the gripper can close, what joint limits it is about to bump. Without that, a world model can imagine a future it cannot actually reach, and plan a motion that is physically impossible.

NVIDIA and 爻方 are answering that gap in opposite ways, and the split became public within five months of each other.

In February 2026, NVIDIA's robotics team shipped DreamZero, a foundation model built on top of GR00T 2, the company's general-purpose robot foundation. Per qbitai's coverage of 爻方's launch, DreamZero takes the joint and gripper state of the robot as just another input signal: a stream of numbers the world model reads, the same way it reads camera pixels. The body is a sensor, not a test.

爻方's iFLYTEK-Embodied-Omni paper does the inverse. It makes inverse kinematics, the math of working backward from "I want my gripper there" to "here are the joint angles," a required answer the model has to produce. The body is not something the model reads. It is a problem the model has to solve on every step, and getting it wrong is a training penalty.

The reader-facing version of the disagreement: should a robot be told where its hand is, or be required to figure it out?

There is a second argument folded into the launch. Pan Jia, in his WAIC remarks as reported by qbitai, is skeptical of the "ship-then-collect" data playbook that has dominated humanoid-robot strategy. The story goes: build a thousand units, deploy them in homes and factories, harvest the teleoperation data, and the next model will be better than the last. Tesla's Optimus program and several Chinese peers are betting on that flywheel. Pan Jia's counter is that the data flywheel only turns if the deployed system is already good enough to be useful. A robot that fails two-thirds of the time in a real kitchen does not produce useful data. It produces a thousand messy teleop logs that all look the same.

That critique is contested, not settled. The Optimus camp would answer that early-generation autonomous vehicles also failed most of the time, and the data still compounded. Whether 爻方's own demos survive contact with real kitchens is the falsifier the company has not yet produced.

A 财联社 (CLS) deep-read on WAIC 2026 framed the moment as embodied AI "leaving the showcase era." That is a polite way of saying what the consumer has already noticed: the robots on stage do not match the robots in the wild. The architectural split between NVIDIA's input-signal fork and 爻方's training-task fork is now a public, named disagreement with money on both sides. Embodied-AI capital in 2026 and 2027 will flow to whichever lab produces the first set of public benchmarks that survives the next round of skeptical third-party evaluation.

The iFLYTEK-Embodied-Omni paper is on arXiv now. The DreamZero code is on NVIDIA's developer site. Anyone with a GPU and a robot arm can run the head-to-head. The result of that race will decide whether Pan Jia's diagnosis, that today's robot brains are missing a layer of self-knowledge, was a sharp critique or a marketing line that happened to be timed for a conference.