On Thursday, US robotics startup Figure introduced Helix 2.5, which it calls the most advanced neural network it has ever built. The humanoid robot walked into 30 homes it had never seen before, used the beds, couches and folding surfaces already in those homes, and got to work — tidying toys, folding towels and making beds — without a single hour of new training or data collection in any of those houses. If that sentence sounds casually insane to you, you're not wrong. Let me break down what actually happened, because the details matter.
01.What exactly is Figure Helix 2.5?
Figure is one of the best-funded humanoid robot companies on the planet, and Helix is the "brain" — the neural network that controls its robot's body. Earlier versions of Helix could do impressive things (Helix 02 once ran a logistics task autonomously for 200 hours), but they shared one weakness: the robot learned from data collected in the specific places where it would work. Take it somewhere new, and it needed fresh training.
Helix 2.5 attacks exactly that problem. CEO Brett Adcock put it plainly in his announcement: "The holy grail for robotics is being able to generalize: doing work in unseen places." The company rented 30 homes in the Bay Area, walked the robot in, and asked it to work with whatever furniture and objects it found — no fine-tuning, no adaptation, no local data collection. Figure says this is the first demonstration of zero-shot whole-body generalization at this scope on a humanoid.
02.The 30-home test: how it actually worked
The robot performed three whole-body behaviours: tidying living rooms (picking up all 13–15 scattered toys and placing them in a basket), folding towels, and making beds (pillows and comforter corners at the top, comforter pulled smooth). Success meant finishing the entire task — no partial credit. Any safety intervention by a human counted as a failed trial.
Here's the part I found most interesting: a single fixed model checkpoint was used across all 30 homes. The toys, towels and bedding used in the evaluations were set aside beforehand and verified (first by an AI model, then by human review) to make sure they never appeared in the training data. The robot was genuinely seeing everything for the first time.
03.Index: the crowdsourced data engine behind it
The secret sauce is a dataset called Index, which Figure launched in late August 2026 after rebranding a data-collection app. Ordinary people around the world record videos of themselves doing everyday tasks, and Figure pays them for it. The numbers are wild: 264,000 app downloads across 108 countries, more than 44,000 weekly active users, over 16 million videos uploaded, and $15 million already paid out to contributors. That works out to roughly 35 minutes of new human-experience video arriving every second.
In a controlled comparison — identical task data, identical architecture, the only variable being Index pretraining — the policy trained from scratch succeeded on 9% of zero-shot trials, while the Index-pretrained policy succeeded on 56%. More than a 6x jump from pretraining alone. The task-level split was also published: bed making came in at 67%, towel folding at 62%, and toy tidying at 40%.
04.Why 56% is a bigger deal than it sounds
Honestly, what surprised me most wasn't the success rate — it's the shape of the result. Helix 2.5 matched the success rate of the older Helix 02 on the same task while using half as much task-specific data, and it did so across 30 environments instead of the one environment the older model was trained in. Figure also reports the first human-to-robot transfer scaling law measured on a humanoid: every time the Index data doubled, robot-action prediction loss fell predictably — so predictably that smaller training runs forecast the largest run's test loss to four decimal places before it began.
Now, the caveats. This was a company-run evaluation, not an independent one. And 56% still means roughly 44% of trials went unfinished. A robot that makes your bed two-thirds of the time is a marvel of engineering but not yet a household staff member. Figure itself says general humanoid robotics remains unsolved.
05.What this means for India
Two things. First, the Index model is quietly creating a new kind of gig economy — people in 108 countries are already earning money by recording videos of everyday tasks. There's no reason Indian users couldn't participate as the platform scales. Second, the trend behind Helix (foundation models that generalize from human behaviour) is the same trend powering AI agents in software — DeepMind alumni just raised $700 million for world models on this exact thesis, something we covered in detail earlier. And if you want to understand how to direct software-based AI agents yourself, our GPT-6 Astra — Zero to Hero Guide has a full chapter on building agents, which is the skill that transfers across this whole wave.
The humanoid gold rush keeps accelerating — D-Robotics just raised $400 million for humanoid chips.
Key Takeaways
- Helix 2.5 performed three whole-body household tasks in 30 unseen homes with no new training — a first at this scope for a humanoid
- Index pretraining drove the jump from 9% to 56% zero-shot success; bed making hit 67%
- Helix 2.5 needed half the task-specific data of Helix 02 while covering 30x more environments
- Index contributors have earned $15 million; Figure has committed $3.5 billion in compute to Helix
- Company-run evaluation, 44% of trials unfinished — impressive, but not a finished product
Frequently Asked Questions
What did Figure's Helix 2.5 robot do?
The Helix 2.5 neural network let Figure's humanoid robot tidy living rooms, fold towels and make beds in 30 rented Bay Area homes — with zero data collected or training done in any of them.
How much did pretraining improve performance?
Pretraining on Figure's crowdsourced Index dataset lifted zero-shot success from 9% to 56% in a controlled comparison.
Are useful home robots here already?
Not quite. It was a company-run evaluation, and 44% of trials still ended unfinished, so robots in every home are not around the corner.
What This Means for You
If you're a student or professional watching robotics from India, this is the moment the field shifted from "robots that memorize one room" to "robots that learn from humans at scale." The data-engine insight — that watching humans do chores can teach machines — will shape the next decade, and it's the same playbook AI agents use in software today. If you want to be on the right side of that shift, start with the software you can touch right now: learn to direct AI agents properly, and cost-control them. Our GPT-6 Astra — Zero to Hero Guide (Digital edition ₹149) walks you from your first prompt to your first working agent in 30 days. The robot that tidies your living room is coming eventually — the skills to direct its software cousins are available today.