Issue #051 — This Week's View

# The Fly That Walks Better Than Your Humanoid's Foundation Model

Here is what I believe: the most important humanoid robotics paper published this week cost nothing, used no data moat, and trained no foundation model — and it should make every humanoid investor recalibrate what they are actually paying for. A new arXiv study ("Humanoid Locomotion with a Fly-Inspired Recurrent Controller," arXiv:2609.27001) shows a recurrent controller with just 3,609 neural states, inspired by insect circuitry, driving a simulated Unitree G1 to complete 61 of 63 locomotion conditions — one short of the privileged reference's 62/63. A controller the size of a rounding error in a VLA model nearly matched the ceiling. The lesson for the humanoid bubble is not that foundation models are useless. It is that we have been conflating two different products — competence and generality — and paying foundation-model prices for what is, in locomotion, mostly a solved low-dimensional problem.

Competence is cheap. Generality is expensive. We keep funding the expensive part to do the cheap part.

Walk through what the fly-inspired study actually found. The controller's deployed locomotion was carried almost entirely by two things: direct body-and-command input, and its own recurrent motor state. Reset that recurrent state before each policy call and success collapsed from 19/21 to 0/21 at nominal yaw. Meanwhile, substituting depth inputs and upstream states at 252 recorded states left actions unchanged, with zero measured descending output throughout intact rollouts. Translation: the sophisticated sensory apparatus contributed roughly nothing to walking. The competence lived in a small, self-sustaining dynamical loop — exactly the architecture biology converged on in insects hundreds of millions of years ago.

Now hold that against where the capital has gone. Per Crunchbase, humanoid startups have raised tens of billions of dollars globally in the last three years alone — Figure AI's reported $39B valuation round earlier this year, Unitree's IPO-track filings, UBTECH's repeated placements, Agibot's and Galbot's nine-figure rounds. The Maze Intelligence catalog tracks 600+ robot models across roughly 350 companies, and the overwhelming majority of humanoid pitches rest on the same syllogism: humanoids need general intelligence; general intelligence comes from foundation models; therefore foundation-model-heavy stacks are the moat.

The fly paper is a direct empirical challenge to the first premise — at least for locomotion, the capability every humanoid demo leads with. If a 3,609-state recurrent loop plus proprioception gets you 61/63 terrain-speed-yaw conditions, then the marginal value of a 7B-parameter vision-language backbone in the walking loop is approximately zero. The honest framing — industry consensus, and our view — is that locomotion is becoming a commodity subsystem, like motor controllers or IMUs. Commodities do not support $39B valuations.

Argument two: the demo layer keeps proving the commodity thesis

The same week, a second arXiv paper (2609.27003) showed an end-to-end pipeline turning monocular runway videos into deployable expressive locomotion on a Booster K1 humanoid — every physical trial completed without a fall, reproducing narrow catwalk foot placement with coordinated torso and arm motion. This is charming, but read it as an investor: expressive, stylistic whole-body locomotion learned from a single smartphone video, deployed reliably. That is not a research moonshot anymore; that is a workflow. When catwalk gaits are a pipeline and terrain walking is a small recurrent net, the "hard" layer everyone demos is racing toward zero margin.

Argument three: the moat is migrating to where the fly paper points

Notice what the fly study did NOT solve and explicitly flags for future work: circuit structure and control resources for higher cognition. The descending pathways were silent. Manipulation, language grounding, long-horizon autonomy — that is where the hard problems actually live, and it is where the backdoor-security paper on FANUC and xArm arms (arXiv:2609.26868) reminds us that even near-term VLA deployment carries unresolved attack surfaces. The rational capital allocation is inverted from current practice: spend commodity prices on locomotion (buy the recurrent controller, license the gait stack), and concentrate bets on manipulation data, safety, and vertical deployment. Companies priced as locomotion stories are mispriced.

The steelman

The strongest counterargument: locomotion is a false proxy, and everyone serious already knows it. Foundation-model bulls will say the valuation case was never about walking — it is about the flywheel of teleoperation data, cross-embodiment transfer, and eventual general manipulation, where no small recurrent controller exists. They will add that the fly result is simulation-only, on a fixed checkpoint, with a survival criterion — and that sim-to-real gaps have humbled better-looking results (the Spiderbot hexapod team, to their credit, notes its 90%+ energy savings and sub-$400 BOM came only after validated sim-to-real transfer). Fair. But this rebuttal concedes my actual point. If locomotion was never the value, then the demo videos — terrain walking, runway struts — that anchor every pitch deck and every billion-dollar round are theater. You cannot sell generality while marketing competence. Pick one, and price accordingly.

What I'd do with a checkbook

Three moves. First, underwrite humanoid rounds on manipulation-data economics, not gait demos; discount any pitch whose centerpiece video is locomotion. Second, watch for locomotion-stack unbundling — the winners here are the Unitrees selling bodies and the open-source controllers, not the integrated-valuation stories. Roughly half the 600+ models in our catalog are locomotion-first platforms (our estimate), and they are competing each other toward hardware margins. Third, treat the fly paper's methodology — circuit-level attribution of what actually drives behavior — as the diligence template. Ask every humanoid team: which parameters carry your deployed behavior? If the answer is "the whole model, we think," you are not buying a moat. You are buying a fly that hasn't learned to land.

The insects solved walking cheaply. The market hasn't yet noticed it was buying the expensive version.

*(Ken Zhang, Maze Intelligence)*