Five robot brains in eight days — and the money follows
Two things happened to Chinese robotics in the same two weeks: the body makers shipped model after model, and the funding data caught up with why.
Three Chinese robot makers released five embodied-AI foundation models in eight days this month — Unitree, AgiBot and Noetix — a pace that says the hardware race is over and the brain is the product.
The sequence, per Red Star Capital Bureau: Noetix launched its Scalabot embodied-AI brand and the first HERON world model on September 8, then HERON-CRA on September 15 — memory past one minute and a 97.8% sock-folding success rate. AgiBot shipped two models in a single day on September 9: AGILE 2.0, which couples perception and control so the robot watches and moves at once, and GE-Act 2.0, a native world-action model built to test whether capability keeps improving as training data grows 100-fold. Then Unitree open-sourced UnifoLM-WLA-1.0 on September 10.
The most concrete artifact of the three is Unitree's. UnifoLM-WLA-1.0 is a 6B-parameter general humanoid foundation model trained on roughly 2,500 hours of real-robot data, and a single set of weights covers 64 tasks spanning tabletop manipulation and whole-body locomotion, driving both two-finger grippers and multi-finger dexterous hands. Its notable design choice splits the action space into end-effector poses, end-effector joints and lower-limb joints, so the policy does not depend on one mechanical chassis. Unitree has published two post-trained checkpoints and a challenge dataset on Hugging Face; the base weights and the training code are still listed as not yet released — worth remembering before calling the release complete.
Noetix is the one to watch on memory. Its CRA model is built for remembering and learning from mistakes across robot bodies, and it reports 97.8% success folding socks — a household task chosen precisely because it defeats most manipulation policies. AgiBot's GE-Act 2.0 makes the scaling argument explicit: manipulation skills that never appeared at 300 hours of data, including folding towels and nesting paper cups, only emerged once training data reached 30,000 hours.
Why it matters: the industry's real bottleneck is generalization, not hardware. Unitree's founder Wang Xingxing said as much at the World Robot Conference — the world's biggest problem is that embodied intelligence does not generalize well enough, and Unitree's largest investment is now AI models. In an eight-day window, three body makers each attacked that bottleneck from a different angle with published numbers.
Chinese embodied-AI funding has flipped decisively from bodies to brains: of roughly 43.8 billion yuan raised in the first half of 2026, brain-focused companies took 22.25 billion yuan, or 50.8%, against 5.6 billion yuan — 12.8% — for robot-body makers.
The concentration is steeper than the split. The top five brain companies alone raised about 17.1 billion yuan, more than 70% of everything the brain category drew, so more than half the sector's capital is chasing a handful of names. Analysts quoted by Red Star Capital Bureau argue the logic has inverted: hardware bodies are no longer scarce, and a company that only integrates parts has too low a barrier to raise independently. One industry executive put it in customer terms — buyers no longer ask whether a robot can do a demo, they sit with a calculator and work out the return on investment.
Why it matters: this tells you what the market thinks the scarce layer is. It also sets up the correction. Valuations on robot brains are being priced like high-margin software, anchored to data throughput and marquee customers, and that holds only as long as the buyers keep valuing the model layer above the machines doing the work. If embodied models commoditize the way language models did, the money is in the wrong half of the stack. The tell is whether the September model wave converts into deployed hours, not benchmark wins.
DiffuSpace, a spinoff of the University of Hong Kong's NLP lab, is training a larger diffusion language model and plans to release it as open source in October.
The team is not new to the architecture. Its Dream 7B model topped the open diffusion-language-model field when it shipped and matched the planning ability of DeepSeek V3, a model with roughly 96 times as many parameters; it has now passed 2.5 million downloads on Hugging Face. Diffusion language models generate text in parallel rather than one token at a time, which makes them dramatically faster on latency-sensitive work — the company says up to 10 times faster than autoregressive models in some scenarios, a claim that depends heavily on the workload and is not independently verified.
DiffuSpace has already put the speed to commercial use: a partnership with agent-computing platform Acrab measured up to a 5-fold speedup for on-device agents, targeting AI PCs, cars, robots and smart-home devices. That is the compelling use case — the sub-second work an agent does constantly, where the quality bar is low and the cost of each round trip compounds.
Why it matters: the competition is no longer theoretical. Google, Inception and Ant Group are all shipping diffusion language models this year, and Inception's Mercury 2.5 is already carrying production voice and coding traffic. The open question is whether dLLM quality scales past mid-size models or whether diffusion stays the cheap, fast supporting cast. October's release gives a real answer: if a bigger open model lands while holding the latency advantage, the case for autoregressive-only stacks weakens. If it ships as a paper without weights, the speed story stays somebody else's product demo.
What to watch: whether Unitree's WLA base weights and training code actually land, and whether DiffuSpace's October release is weights or a writeup. The claims in this wave are only as good as the artifacts.
If the robot brain is where the value is, should the body makers keep building their own — or license one from the lab that already has the data? Tell us in the comments.
Sources: Unitree (GitHub)
