Figure Helix 2.5: zero-shot housework in 30 never-before-seen homes — with no data from these households, no fine-tuning
🔎 No more lab demos: robotics faces the real-world test
Humanoid robot demos share a recurring flaw: they always run in the same hallway, in front of the same shelves, under the same lighting. Change a single parameter and the magic video turns into a weeks-long retraining campaign. Generalizing to the unknown — the real test — remained the industry's blind spot.
On September 17, 2026, Figure published an announcement that directly tackles this weak point: Helix 2.5, its AI model, evaluated in real-world conditions across 30 rented homes in the Bay Area. With no data collected from these households. No fine-tuning on their layouts, their furniture, or their objects. A single model, frozen, identical everywhere.
Why does this matter now? Because the race for the home humanoid is entering a new phase: after proof of endurance in factories, the battle is shifting to multi-environment generalization. Here's what the announcement really says, what it leaves out, and where Figure stands against the competition.
The essentials
- 30 never-before-seen homes: Figure rented 30 homes in the Bay Area and evaluated Helix 2.5 there with no data collected in those households, no fine-tuning on their environments or objects, and a single frozen checkpoint (official announcement).
- One model, three whole-body behaviors: tidying living rooms (13 to 15 scattered toys to put in a basket), folding towels, making beds.
- Pre-training on Index changes everything: zero-shot success goes from 9% (policy trained from scratch) to 56% — more than 6x, with data, architecture, and evaluation otherwise identical.
- Massive adaptation savings: Helix 2.5 matches a policy trained with data collected on-site while using half as much adaptation data, and generalizes to 30 unseen homes ("behavior specification 2x cheaper while its scope expanded 30x").
- A claimed scaling law: the first human→robot transfer scaling law measured on a humanoid; the action prediction loss drops predictably with each doubling of Index (the forecast for the largest run was accurate to within 0.54%).
- Acknowledged limitations: no released weights, no API, no pricing, no date — a research milestone, not a product.
Tools: Humanoid AI systems in the race (September 2026)
The comparison below lists the robotic AI systems whose approaches directly compete on the path to the "generic brain". None of them is sold to the general public; prices refer to undisclosed deployment models.
| System | Main use | Price | Ideal for |
|---|---|---|---|
| Helix 2.5 (Figure) | Whole-body household chores, zero-shot generalization across multiple homes | Not disclosed (September 2026) | Research; milestone for domestic generalization |
| Helix 02 (Figure) | Full 8-hour factory shifts, autonomous logistics (200 h) | Not disclosed (September 2026) | Industrial deployment in controlled environments |
| Skild S1 (Skild AI) | Learning a task from a single video, without fine-tuning | Not disclosed (September 2026) | Express adaptation to a specific task |
| OM-1 | Training on human demonstrations | Not disclosed (September 2026) | Demo → autonomous execution pipelines |
| Digit 5 (Agility Robotics) | Humanoid designed from the ground up for work | Not disclosed (September 2026) | Structured professional environments |
Keep in mind that these systems are not measuring the same thing, on the same tasks, under the same conditions. That is precisely the purpose of this article.
What Figure announced on September 17, 2026 — A direct answer
Figure rented 30 real homes in the Bay Area and sent its humanoids to work there without any data from these households: no prior collection, no fine-tuning, no adaptation to the layouts, furniture, or objects. That is the exact meaning of the "zero-shot" claim made in the official announcement.
The test covers three whole-body behaviors — not just arms. Tidying up a living room, with 13 to 15 scattered toys to put back into a basket. Folding towels. Making beds. A single foundation model produces all three behaviors.
The success criteria are unusually rigorous for the industry. A task only counts as successful if it is fully completed: every toy put away, every towel folded, every bed made from start to finish. No partial credit. And any human safety intervention counts as a failure.
Another detail that matters: the objects used for the evaluation — toys, towels, bedding — did not appear in the task specification data, a point verified first by an AI model, then by human review. Along the way, Figure also claims a first: the largest zero-shot whole-body generalization demonstration ever performed on a humanoid.
Brett Adcock, founder and CEO of Figure, sums up the intent: "The holy grail of robotics is the ability to generalize: to work in places never seen before. We rented 30 homes in the Bay Area and are performing tasks in them with no new training whatsoever. Helix 2.5 was built to answer a harder question: can a humanoid walk into a home it has never seen and immediately get to work, with its whole body, entirely on its own?" (translation, The AI Insider).
My take: the methodology — rented homes, frozen checkpoint, strict criteria with no partial credit — is more convincing than 90% of the robotics demos published over the past three years. The devil, as always, is in the details of the protocol. More on that below.
"Zero-shot": what the word covers — and what it doesn't — Direct answer
Helix 2.5's "zero-shot" applies to environments and objects, not to the behaviors themselves. The three tasks remain specified through fine-tuning, whose data was collected elsewhere — it's Unite.AI that correctly points this out in the very first lines of its analysis.
In other words: Figure didn't teach the robot to make these beds, in these homes. It taught the robot to make beds somewhere else, then evaluated it in 30 completely unfamiliar households. The nuance is essential — and, to its credit, Figure doesn't hide it.
Three safeguards nonetheless give the result real weight. First, the evaluation objects did not appear in the task specification data, verified first by an AI model and then by humans. Second, no evaluation task accounts for more than 1.90% of the Index dataset — so there is no dominant task on which the system could have quietly memorized the answer. Third, a single, frozen checkpoint was used for all 30 homes.
The right reading: Helix 2.5 demonstrates massive environment transfer, exactly the kind of generalization that domestic robotics was most lacking. The wrong reading — the one behind the press headlines — would be "the robot can do anything in any home." Neither Figure nor the serious sources claim that.
The numbers: from 9% to 56%, and a scaling law — Direct answer
Pre-training on Index raises the zero-shot success rate from 9% to 56%: a gain of 47 points, roughly 6.2x the performance of a policy trained from scratch.
| Configuration | Zero-shot success rate (complete tasks) |
|---|---|
| Policy trained from scratch (identical data, architecture, optimization, hyperparameters, and evaluation) | 9% |
| Helix 2.5, pre-trained on Index | 56% |
The protocol is clean: everything is held constant, and only one variable changes — pre-training on Index or its absence. It's rare to see such a crisp ablation in a public announcement.
Second result, and not the least: Helix 2.5 matches the success rate of a Helix 02 policy trained on data collected directly in the tested environment. With half the adaptation data, and while generalizing to 30 unseen homes. Figure sums it up in its own words: "behavior specification 2x cheaper while its scope expanded 30x." A word of caution, however: AlphaSignal notes that these two dimensions — data halved, reach multiplied by 30 — don't combine into a single quality multiplier. Beware of arithmetic shortcuts.
Third element, potentially the most important: the scaling law. Four models were trained on nested subsets of Index, covering a range of 8x in pre-training data volume. With each doubling, the action-prediction loss drops in a predictable way — to the point that Figure claims to have predicted the test loss of the largest run before even training it, with an error of 0.54% of the variation measured across the entire range. If this behavior holds up under independent replication, it may prove more decisive than the 56% itself: an isolated milestone becomes an industrial trajectory.
The infrastructure behind these numbers is dizzying: Index generates roughly 35 minutes of new human experience data per second, and Figure has committed $3.5 billion in compute resources to train Helix (The AI Insider).
What third parties still need to establish, AlphaSignal lists: broader task coverage, per-trial reporting, independent replication, and a demonstrated link between action-prediction loss and real-world completion rates. KiaDev adds a sound methodological caveat: the 56% is a reported result, backed by an evaluation defined by Figure itself — without full benchmark specification, without a task-by-task breakdown, without comparison against other neural networks, and without a published number of trials.
From Helix 02 to Helix 2.5: the factory first, the living room next
Helix 02 had proven its endurance: full 8-hour shifts in a factory, without human intervention, from unloading dishwashers to a 200-hour autonomous logistics task (our dedicated article). But its learning relied on data collected where the robots operate — in the deployment environments themselves.
This is precisely the dependency that Helix 2.5 removes. As KiaDev puts it, a system that requires fresh data for each environment has a radically different deployment profile from a system that can be evaluated without local data collection. Concretely: moving from a new factory site to a new home no longer requires an on-the-ground data collection campaign, just a pre-trained checkpoint and a task specification. That's the entire economic question this result addresses.
The most telling qualitative improvement concerns in-task self-correction, which Figure attributes to Index pre-training: stepping back to reposition itself, shifting its support, walking all the way around a bed to fix a fold (Unite.AI). Gulf News emphasizes this point in its explainer: no map, no task-specific training, no prior data — the robot arrives, gets to work, and recovers from its mistakes without direct human intervention.
As a reminder, Helix was designed from the outset to bring humanoids into homes (our analysis of the initial launch). Helix 2.5 is the first credible answer to the problem that launch posed: how to serve millions of homes, all different from one another, without scanning each one individually?
Index: Figure's real moat isn't the robot, it's human data
What sets Figure apart isn't found in its actuators, but in Index, its dataset of human behaviors at global scale. Helix 2.5 is presented by the company as "the most advanced neural network Figure has built" — and everything rests on this pre-training.
The scale figures set the tone: roughly 35 minutes of new human experience data produced per second, continuously. The partnership with Brookfield, which makes more than 100,000 residential units available worldwide, aims to make Index the largest humanoid pre-training dataset in existence (Figure, Project Go-Big).
First published result of this strategy: zero-shot transfer from 100% egocentric human video to robot navigation in cluttered spaces, on natural language commands ("go to the fridge"), without any robot demonstrations — a claimed first in the field.
A quick naming aside: while PageIndex bets, on the RAG side, on structured indexing rather than raw vector similarity, Figure's Index structures human behavior on the robotics side. Two different domains, the same intuition: organized data beats naive search.
My take: this is where the real race is being played out. Compute can be bought — $3.5 billion proves it. Egocentric human experience data, on the other hand, can't be downloaded. Whoever holds human behavior at scale holds the generic brain of robotics, and Brookfield just showed that Figure understood this before everyone else.
The race for the generic brain: Skild, Spirit AI, OM-1, Agility
Figure is not alone in the field, and domestic zero-shot is clearly becoming the new front line of robotics.
Skild AI showed off Skild S1, a robot learning a 10-minute task from a single video: 11 minutes elapsed between the demo and autonomous execution, with no fine-tuning at all (our article). Where Figure generalizes across environments, Skild generalizes task acquisition. Two different axes of the same ambition.
Spirit AI, for its part, claims a 90% success rate on simple tasks in a structured environment. An impressive figure on paper — but compared to Figure's 56% on complete household tasks across 30 unknown homes, it measures something else. A structured, known environment has absolutely nothing to do with a living room never seen before, lit differently, and filled with unfamiliar objects. OM-1, meanwhile, trains on human demonstrations — a complementary approach, oriented toward a demo-to-execution pipeline.
And the hardware is following suit: Agility Robotics has unveiled Digit 5, presented as the first humanoid designed from the ground up to work (our coverage). The brains are progressing, and so are the bodies.
My take: be wary of comparison charts lining up isolated percentages. The metric that matters in 2026 is no longer "it works in our lab," it's "it works in a place the system has never seen." On this specific — and published — metric, Helix 2.5 is today the most ambitious benchmark.
A research milestone, not a product — Direct answer
Helix 2.5 is neither for sale nor accessible: no published weights, no API, no price, no date. Gulf News says it plainly: it's a research milestone, not a consumer robot. Any announcement of an "available household robot" would be a fabrication.
What's still missing to make the leap? A lot. Task coverage (three behaviors — that's demonstrative, not a catalog). Reliability — 56% of tasks completed is a scientific milestone, not a commercial standard, where we talk about 99% and above. Safety in unstructured environments, with children, animals, stairs. And the cost of domestic humanoid hardware, whose final price nobody knows.
The path, however, is taking shape. The Brookfield partnership covering more than 100,000 residential units shows that data — the fuel of generalization — is being built right now, massively, in real homes. Helix 02 proved the factory, Helix 2.5 proves the home; the product remains to be proven.
My take: the real value of this announcement is methodological, not commercial. Figure isn't selling anything today — it's demonstrating that the recipe of massive pre-training + lightweight adaptation works on a physical body, in the mess of the real world. That's what the competition will have to replicate, numbers in hand.
❌ Common Mistakes
Mistake 1: Believing that "zero-shot" means "zero training"
False. The behaviors (toys, towels, beds) are always learned through fine-tuning on data collected elsewhere; zero-shot applies to the evaluation environments and objects. The fix: read Figure's announcement in full rather than the headlines — the company itself frames its claim.
Mistake 2: Reading 56% as a "commercial readiness rate"
This is not a product reliability label. It is the rate of fully completed tasks under strict criteria (no partial credit, human intervention = failure), on a benchmark defined by Figure, with no independent replication to date. The fix: wait for per-trial reporting and third-party replications before any extrapolation.
Mistake 3: Comparing 56% and 90% as if they measured the same thing
Spirit AI reports 90% success on simple tasks in structured environments; Figure measures 56% on complete household tasks in homes never seen before. The tasks, the criteria, and the environments differ at every term of the comparison. The fix: compare the protocols before the percentages.
Mistake 4: Forgetting the data in the equation
Commenting on "the robot" while overlooking Index — 35 minutes of human data per second, $3.5 billion in compute, no equivalent real-estate partnership among competitors — leads to misjudging the reproducibility of the result. The fix: evaluate the data asset as much as the model.
❓ Frequently Asked Questions
Is Helix 2.5 available? How much does it cost?
No. Helix 2.5 is a research milestone: no published weights, no API, no pricing, no commercialization date to this day (September 2026). Figure only communicates internal evaluation results. The only reliable source to follow what comes next remains Figure's official blog, where the methodology and results are detailed.
What exactly does the 56% success rate mean?
It's the rate of tasks fully completed — zero partial credit — across 30 never-before-seen homes, with any human safety intervention counted as a failure. With otherwise identical data, architecture, and evaluation, a policy without Index pre-training only reaches 9%. A result reported by Figure, not yet replicated by third parties.
What's the difference from Helix 02?
Helix 02 masters whole-body control over long durations — 8-hour factory shifts, a 200-hour autonomous logistics task — but learns from data collected in the very environments where the robots operate. Helix 2.5 adds generalization: evaluation in 30 homes without any local data or on-site fine-tuning.
How does the robot learn its behaviors?
Through fine-tuning, on data collected outside the tested homes, after massive pre-training on Index, Figure's dataset of human behaviors (around 35 minutes of new data per second). Zero-shot covers the evaluation environments and objects, not the specification of the behaviors themselves.
When will we see these robots in our homes?
No date exists. The path is taking shape: Brookfield is making more than 100,000 residential units available to build up the data, but the 30 tested homes remain a controlled experiment. Between 56% success on three behaviors and a reliable home product, there remains task coverage, reliability, and cost to conquer.
✅ Conclusion
With a single announcement, Figure turned multi-environment generalization from a roboticist's dream into a measurable metric — and the jump from 9% to 56% thanks to Index says one thing above all: the race for home robotics is now, first and foremost, a race for human data. What's next? Follow Figure's official announcements — and keep an eye on who, among the competition, will be the first to publish a comparable figure in a never-before-seen home.