Issue #054 — This Week's View

# Human Data Is Eating the Teleop Moat

Maze Intelligence | VIEWPOINT | Ken ZHANG

Here is what I believe: the teleoperation data moat — the single most-cited defensibility argument in humanoid investing over the past two years — is already collapsing, and the papers证明 it are not coming from well-funded labs with thousand-robot fleets. They are coming from academic teams with a single Unitree G1, a mocap rig, and a new recipe: train on human data directly, skip the robot entirely.

My evidence is Workhorse, a system published on arXiv this month (arXiv:2610.09117) that learns contact-rich whole-body loco-manipulation from *robot-free human demonstrations*. A visual planner predicts five body targets — torso, both wrists, both feet — and a reinforcement-learning tracker executes them on the robot. Both policies are trained separately on the *same recorded human poses*, with no retargeting step. On a real Unitree G1, it sorts boxes using its hands and a kick, catches a thrown box, and climbs a suitcase. Push the robot mid-task and it recovers; steal its box and it adapts.

If you have underwritten a humanoid company on the premise that "our teleop fleet generates proprietary data nobody else has," you should read that abstract twice.

Three reasons the moat is evaporating

One: the retargeting bottleneck was the moat, and it's gone. The old argument went: human motion doesn't map cleanly to robot morphology, so whoever builds the best retargeting + teleop pipeline + fleet accumulates a data asset competitors can't replicate. Workhorse sidesteps retargeting entirely — both policies train directly on recorded human poses. The scarce asset shifts from *operated robot hours* to *human motion capture*, which is dramatically cheaper and, crucially, not exclusive to anyone. Your competitor can rent mocap data. Roughly speaking, that collapses the moat from a capital moat to a data-sourcing moat — and data sourcing is a race to the bottom. Two: the numbers are honest, and they're still good enough to matter. Let me be precise about what Workhorse actually claims, because the details matter more than the demo reel: 77% task completion in simulation on box sorting, 64% under perturbation forces of 40 N·s, and 83% for a second humanoid retrained from the same demonstrations — no pushes. These are not production numbers, and I'm labeling them as such: they are the authors' own reported simulation results. But here is the point VCs should internalize: *zero teleoperation hours* produced a system that recovers from human pushes on a real G1. The baseline for "what you get without a teleop fleet" just moved. Three: cheap sensing is doing the same thing on the manipulation side. The same arXiv batch includes SAFE (arXiv:2610.08802): two passive PVDF acoustic sensors on a compliant Fin Ray gripper, a classifier, and a 104 Hz closed-loop controller on a Jetson Orin Nano achieving 91.3% success across 46 closed-loop trials — and 82.4% on novel objects, per the authors' reported results. Fixed-force baselines drop to 0% on mismatched object conditions. In other words, the *tactile* half of manipulation — historically another justification for proprietary fleets and custom hardware — is being solved with a few dollars of sensors and a HistGradientBoosting classifier. The commons is eating the moats from both ends.

The pattern investors should price in

Across my catalog work at Maze Intelligence — we track 600+ robot models across roughly 350 companies globally, by our count — the teams raising on "data moats" overwhelmingly mean teleop fleets, proprietary teleop pipelines, or in-house task suites. The academic frontier is publishing recipes that need none of these. Workhorse-style human-data training will not stay in academia; the recipe is legible, the hardware platform (Unitree G1) is commodity, and the incentive to productize is enormous.

My view, and I'll label it as our editorial judgment rather than data: within 18–24 months, "we have a teleop fleet" will shift from a moat claim to a *cost* claim — and VCs will start asking why a company is paying operators when it could rent motion capture and train from human data. The companies that survive this shift will be the ones whose defensibility lives elsewhere: deployment relationships, service economics, reliability track records, vertical integration into specific factories. The ones whose entire pitch was "our data" should be re-underwritten today.

The steelman

The strongest counterargument deserves a fair hearing. Simulation success rates are not deployment success rates, and Workhorse's 77% (authors' own figure, in a simulated copy of the demonstration room) is a long way from the reliability a warehouse buyer demands. Teleop fleets generate data *in the exact distribution the deployed robot faces* — real lighting, real clutter, real customer workflows — while human-data pipelines still face a sim-to-real and morphology gap that papers like Workhorse narrow but do not close. A fleet is also a *compounding* asset: more robots deployed means more data from the tails that no mocap studio ever records. And SAFE's 82.4% on novel objects is impressive precisely against a low bar — fragile-object handling at commercial reliability likely needs orders of magnitude more real-world contact data than a lab study provides. Fair points, all of them.

But notice what this defense concedes: it defends teleop data as *useful*, not as a *moat*. If the recipe for good-enough loco-manipulation is publishable by an academic team with one robot, then exclusivity — the definition of a moat — is gone, even if teleop data remains a quality edge. Useful data is an input. Moats are structures. Don't let founders sell you the former as the latter.

The bottom line

The humanoid sector does not lack capital — it lacks honest theories of defensibility. Workhorse and SAFE are two data points, not a trend line, but they point the same direction: the input scarcity that justified 2023–2025 humanoid valuations is being engineered away faster than the term sheets assumed. Ask every humanoid founder in your pipeline one question: *if human data plus commodity hardware can produce what you produce, what exactly do you own?*

If the answer is "our teleop fleet," you now know what the academic labs believe about that. So do I.

---

*Maze Intelligence tracks 600+ robot models across ~350 companies globally. Counts and company details are from the Maze/FinBrain catalog; model counts marked "roughly" are our estimates.*