In LatePost's Q2 embodied AI roundup with Alphaist partner Chen Zhe, the durable story is not the Beijing humanoid half marathon or Figure's 200 hour livestream.
If you wanted a clean win for embodied AI in Q2 2026, the Beijing Yizhuang humanoid half-marathon looked like one. On 19 April 2026, more than 100 humanoid robots ran the 21.1-kilometer course alongside human athletes; the winning biped, "Shandian," finished around 50:26, per Xinhua and Beijing Daily reporting. A 200-hour continuous livestream from Figure 03, sorting roughly a quarter-million logistics packages with near-zero failures according to ithome, Huxiu, and PCONLINE, looked like another. The durable story LatePost host and Alphaist partner Chen Zhe spent an entire Q2 roundup arguing for, in episode 170 ("具身季报 26Q2") published 29 June 2026, is not the race result or the livestream. It is the definition war over "world models," a category everyone is funding and almost nobody can pin down in a sentence.
The standard embodied-AI stack has three layers worth distinguishing. A vision-language-action model (VLA) maps camera input and language instructions to robot actions, functioning as a policy. A motion planner turns abstract goals into joint trajectories. Neither by itself is enough for a robot that has to navigate a warehouse aisle or a household mess. A world model, in its loosest working definition, is a learned simulator: a generative model of how a physical scene will respond to an action, trained on video, action, and sometimes proprioception. NVIDIA's Cosmos 3, launched this quarter as an "open frontier foundation model for physical AI," is the most prominently marketed example. The pitch is direct-action generation: skip rendering a video and planning on top, instead predict the consequence of an action and act.
Three circles are converging on that idea. LatePost and Chen Zhe frame them as generation, action, and world, each previously a separate research thread, now folding into a single training stack. Physical Intelligence's π0.7, per the company's own write-up as recounted on the podcast, fuses a VLA with a lightweight world-model head so the policy can reason about its own likely errors. Generalist's GEN-1, the company argues on the same podcast, should be pretrained from scratch on action and video without a pre-imposed route. Both routes are still "podcast plus company blog" claims, and the open disagreement on the term "world model" is not a coincidence. LatePost frames this as definitional mush that has not slowed the money.
The money, in fact, is chasing a thesis at the velocity the LLM wave set in 2023. LatePost describes a venture-capital pattern in which a category-grade name forms, capital moves at LLM-mania pace, and the boundary between "world model," "video generator," and "action policy" stays blurry on purpose. The OpenAI Robotics team was formally announced in Q2, per the podcast; Google DeepMind's ER1.6 was released this quarter, again per the podcast. Both are infrastructure bets whose payoff depends on what a world model can and cannot do once one is pinned down.
The hardware side has not waited. At ICRA 2026 (the IEEE International Conference on Robotics and Automation), Chinese high-DoF dexterous hands drew unusual attention, including the 舞肌 (Wuji) hand; a direct-drive-versus-tendon-drive debate continues, with Optimus, per LatePost, still committed to tendon drive, trading peak force for cleaner backdrivability and packaging. Xingdong Ji-yuan (星动纪元) is reportedly partnering with China Post on to-B humanoid deployments, per LatePost, a small data point that says more about China-side commercial motion than about whether the underlying definition is settled.
In China, the question is sharper than in the United States because the same category names (world model, embodied foundation model, 具身大模型) are being attached to different research programs by competing labs, and the venture capital chasing them is local. The Yizhuang marathon and the Figure livestream both functioned as proof-of-life rituals for a category whose central claim, that a robot can predict a physical world well enough to act inside it on the first try, has not been independently benchmarked at the scale language-model benchmarks reached around GPT-4.
The honest falsifier is straightforward. A world model is real in the way GPT-4 was real for language: when a downstream consumer can pick one up, point it at a new scene, and get a useful prediction back without fine-tuning. Right now the closest public attempts are Cosmos 3, π0.7, and GEN-1, and each makes a different promise about what "useful" means. Until that converges, the safest read of Q2 2026 is what the wire already showed: hardware is racing, money is racing faster, and the underlying category is still in search of a boundary.
The watch item for Q3 is whether any of the three circles (generation, action, or world) produces a benchmark that survives contact with the Figure 03 sortation task or the Yizhuang course. If yes, the chase is the start of a real capability story. If not, "world model" is on track to go the way of "metaverse," a word that lifted a category of money before the science delivered.