Figure Helix 2.5: Zero-Shot Autonomy Across 30 Unseen Homes

Figure released Helix 2.5 on September 18, 2026, and answered the question explainx.ai's teaser analysis posed the day before: yes, the "breakthrough" is locomotion and whole-body generalization learned from human video, not robot teleoperation. To prove it, Figure rented 30 homes across the Bay Area and sent Helix 2.5 into each one with zero data collected there, zero fine-tuning, and zero prior exposure to the objects it would need to handle. It tidied living rooms, folded towels, and made beds. Here's what actually shipped, the ablation that backs up the headline claim, and the reasons to stay skeptical of the promotional footage even while taking the underlying result seriously.

TL;DR

What actually shipped

Helix 02, Figure's prior model, coordinated a humanoid's whole body over long horizons — unloading a dishwasher, running an autonomous logistics task for 200 hours — but it learned from data collected in the specific places it operated. explainx.ai covered its collaborative bedroom-tidying demo back in May. Helix 2.5 targets a harder, more falsifiable question: can the same model enter a home it has never seen and get to work immediately, with no per-location adaptation at all?

Figure's answer: pretrain a single foundation model on Index — its crowdsourced dataset of everyday human activity, which explainx.ai covered at launch in August — then adapt that one base model to three distinct behaviors spanning locomotion, rigid and deformable object manipulation, bimanual coordination, and active perception: tidying a living room, folding towels, and making a bed. All three were then tested across 30 Bay Area homes with zero data collected in any of them, and with evaluation objects verified (by an AI model, then human review) to not appear anywhere in the task-specific training data.

The number that matters more than the video: 9% to 56%

Demo footage is the least trustworthy artifact in robotics — explainx.ai's own checklist for judging robot demos exists because of exactly this problem. What makes Helix 2.5's release different is that Figure ran a controlled ablation specifically to isolate the cause of its zero-shot capability, rather than just showing footage and asserting a mechanism.

The setup: train two policies on identical task-specification data — data that did not include any evaluation home or object — with everything else held fixed: architecture, optimization, hyperparameters, downstream data, evaluation. The only variable that changed was the starting point. One policy initialized from random weights; the other initialized from the Index-pretrained Helix 2.5 base model.

In blind evaluations, the randomly initialized policy succeeded on 9% of zero-shot trials. The Index-pretrained policy succeeded on 56% — over six times higher. Success required completing the entire task (every toy picked up, every towel folded, the whole bed made), with no partial credit given. That's a controlled, isolated variable, not a correlation — which is what makes it a genuinely stronger claim than "the demo looked impressive."

A scaling law for human-to-robot transfer

Figure also makes a second, more speculative claim: that the relationship between pretraining data and downstream robot capability behaves like the scaling laws that transformed language models — where next-token prediction improves predictably with more data and compute, letting researchers forecast large runs from small ones.

To test it, Figure trained four models on nested subsets of Index spanning an 8x increase in pretraining data, holding model size and downstream training fixed, then measured a held-out action-prediction loss. Loss fell predictably with each doubling of Index data — precisely enough, Figure claims, that the smaller runs forecast the largest run's test loss to four decimal places before that run finished training, with forecasting error at just 0.54% of the variation across the full 8x data range.

Two caveats worth keeping in view: this measures data scaling only, with model size and downstream training held fixed, so it doesn't tell you how model size or task-specific fine-tuning data trade off against pretraining scale. And a single scaling curve on one company's internal dataset is not the same evidentiary weight as the years of replicated LLM scaling work it's being compared to. Still, if it holds up under scrutiny, it's the kind of result that changes how confidently a company like Figure can plan its next data and compute spend — which is presumably why Figure is now committing $3.5 billion to training Helix and says Index ingests roughly 35 minutes of new human experience every second.

Data efficiency: half the data, thirty times the reach

A separate result addresses cost rather than raw capability. Figure compared Helix 2.5 against a Helix 02 policy trained on the same task, where Helix 02's training data was collected directly in the environment it was evaluated in. Helix 2.5 matched that policy's success rate using half as much task-specific adaptation data — and did it zero-shot, across 30 homes it had never touched, rather than in the one environment it was trained for.

That's the economic argument underneath the technical one: if pretraining on human video genuinely substitutes for expensive, slow, per-environment robot data collection, the cost of deploying a capability across N new locations stops scaling with N.

The skepticism, and why it's reasonable

Not everyone is buying the presentation. One widely shared reaction to the release video pointed out that the camera cut away from each of the three demonstrated tasks before any of them visibly finished — a real, specific critique, not generic hater noise, and exactly the kind of thing explainx.ai's robotics-demo checklist flags as a red flag: cuts hide failures, resets, and retries, and a single unedited take is worth more than a polished montage.

The honest reconciliation is that Figure's write-up and its promotional video are different artifacts serving different purposes. The evaluation methodology — fixed checkpoints per task across all 30 homes, no weights adapted post-hoc, blind grading with published pass/fail rubrics for toys, towels, and bedding, and explicit "all or nothing" scoring with no partial credit — is meaningfully more rigorous than a highlight reel. The video, like nearly every robotics launch video, is edited for pace and impact rather than methodological transparency. Trust the ablation and the rubric; treat the video the way you'd treat any company's own trailer.

Honest limitations

• This is Figure's own self-reported result, with its own methodology, graders, and success criteria — there's no independent third-party reproduction yet.
• The demo video specifically drew criticism for cutting away before tasks completed, which undercuts the persuasiveness of the footage even if the underlying blind-evaluation numbers are sound.
• The scaling-law claim measures data scaling only, with model size and downstream training fixed — it doesn't establish how those other variables trade off.
• 56% zero-shot success, while a large jump from 9%, is still under half of all-or-nothing task attempts failing — this is evidence of generalization, not evidence the problem is solved. Figure's own post is explicit that "general humanoid robotics is [not] solved."

What this means for builders

Most readers of this site aren't building humanoids, but the underlying strategy is worth studying regardless of category: Figure treated its hardest constraint — the cost and speed of collecting paired robot demonstration data — as a data-sourcing problem, not a modeling problem, and routed around it by paying people to film ordinary chores at scale. The model work followed from that data position, not the other way around. That's the same pattern behind world models that learn environment dynamics from passive video instead of costly interactive rollouts — identify the input your competitors assume has to be expensive, then find the substitute that already exists at scale.

Related on explainx.ai

• Figure teased an AI breakthrough — here's what the evidence pointed to
• Figure Index: the crowdsourced training dataset behind Helix
• Figure Helix 02: two humanoids tidying a bedroom together
• UBTECH's 10,000-unit-per-year humanoid plant in Liuzhou
• Gemini Robotics 2 and whole-body intelligence
• 1X NEO: 25-DOF hands and a physical API
• Mistral Robostral and embodied navigation
• World models explained: learning dynamics from passive video
• Official source: Figure — Helix 2.5: Zero-Shot 30-Home Generalization

This post is sourced to Figure's own September 18, 2026 release write-up and demo video. Success rates, the pretraining ablation, and the scaling-law figures are Figure's self-reported methodology and numbers; no independent third-party reproduction is available at time of writing.