Humanoid Robots

Unitree G1 is the de facto humanoid research and validation platform

Across a week of arXiv Physical AI papers, Unitree's G1 is the hardware on which results are proven: zero-shot whole-body policy transfer, weight-space diffusion policies deployed without fine-tuning, and GPT-6-Astra arm and walking control with Wuji hands. Dyna's CEO credits Unitree with shrinking the sim-to-real gap to a very strong breakthrough. This makes Unitree the Linux-like reference body for the field, while Western names like Figure, Apptronik and Boston Dynamics appear mostly as comparison points. The commercial implication is that research velocity is accruing to a Chinese hardware vendor, not to the best-funded US humanoid makers.

Weekly funding was $2,417M in the week of Aug 24 and $6,222M in the week of Sep 28, but only $10M, $72M and $13M in the three most recent weeks. The recent flow is Ascensia's $36M Series B and a $12.5M seed from a16z Speedrun and NFX, and the totals are dominated by outliers like Project Prometheus. The median deal is now small and early-stage.

Why it matters · Investors should read the headline totals as outlier-driven and treat the underlying cadence of roughly 4 deals a month as the true run rate.

Physical Intelligence's π0.5 is the default base policy across many papers, and its checkpoints were graded inside Ropedia W0's world-model proxy with a 0.984 ranking correlation, showing that world models can evaluate policies without physical access. AgiBot World, Fourier's ActionNet and 1X's world model all feed or benchmark this stack. Research is also contesting the architecture itself, with discrete actions matching or beating continuous ones.

Why it matters · Value is shifting toward whoever owns the datasets, simulators and evaluators, not just the robot body.

Papers argue that Figure, Boston Dynamics and Apptronik over-focus on end-effector dexterity, and that using the whole body as a contact surface is the bigger unlock. Full-body compliance (SoftMimic) beats upper-body-only approaches like GentleHumanoid. Meanwhile generalist policies collapse on multi-step tasks, falling from 85.4% on a single subgoal to 0% at five in simulation.

Why it matters · Demos overstate deployable capability, so the winners will be those who close the multi-step and whole-body gap.

A Chinese podcast guest notes Unitree at about 200B RMB and says the upside from here is a roughly 20x scenario that rests on an unquantifiable market, unlike EV substitution. Separately, Ultra leases warehouse packing robots as a service, with 500,000+ orders packed, betting that non-humanoid robots with recurring revenue scale faster than humanoid demos. AgiBot, by contrast, is described as having ample capital supporting a broad product line.

Why it matters · Capital may rotate toward deployable, revenue-generating form factors unless humanoids show real unit economics.