Watch the video with chapters & transcript →
AI TOOLS
In short: No: Figure's Helix 2.5 robot succeeded in only 56% of attempts across 30 unseen homes, up from 9% before it learned from human videos — a six-fold jump, but still not a product.
Ani Björkström, the Stockholm-based AI-for-finance consultant, breaks down Figure AI's September 2025 Helix 2.5 demonstration — in which a humanoid robot tidied toys, folded towels and made beds in 30 real rented California homes it had never entered before. Rather than focusing on the robot footage that went viral, the video examines the data pipeline behind it, including a phone app that crowdsources housework videos, and asks what the result actually means for investors evaluating Figure's $39 billion valuation.
Figure rented 30 real homes in California and gave its humanoid robot three jobs: clear 13 to 15 toys off the living room floor into a basket, fold towels, and make the bed. The robot had never been inside any of these homes, had never seen the specific toys, towels or bed sheets before, and ran on the same unmodified AI brain in every home.
The scoring was strict: missing even one toy counted as a full failure, with no partial credit. Across all attempts, the robot completed 237 of 420 tries, a 56% success rate, with bed-making as the strongest task at roughly 67%.
Figure trained two versions of the same robot brain under identical conditions and tested both in the 30 homes. The version trained from zero succeeded 9% of the time; the version that had first watched human videos of people doing the same kinds of tasks succeeded 56% of the time.
The human videos came from Figure's Index app, launched in August 2025, which pays ordinary people to film themselves cooking, cleaning and doing laundry. At launch, new video was arriving at a rate of 30 minutes every second — nearly five years of footage per day — and that pace rose to 35 minutes per second a few weeks later. Figure also found a scaling pattern: four robot brains trained on progressively more video, up to eight times more, improved in a predictable line, letting the company forecast the largest brain's performance before training it.
Ani Björkström lists five gaps finance professionals should weigh before reading the result as proof that general robotics is solved. First, 56% success also means 44% failure, and no outside party has yet verified Figure's own test results. Second, the robot was pre-trained on these three specific tasks elsewhere before entering the new homes, so it generalized across locations, not across unfamiliar tasks.
Third, viewers online noted the released footage cuts away before some tasks are shown fully complete, meaning the test numbers should be trusted over the trailer. Fourth, all 30 homes were in California, not in a small Stockholm apartment or a factory setting — Figure itself states plainly that general robotics is not solved. Fifth, the comparison to self-driving cars is a caution: those looked nearly ready more than a decade ago, and the final percentage points took years and billions of dollars more.
| Step / Event | Detail | Result |
|---|---|---|
| Index app launch | August 2025; pays users to film housework | 264,000 downloads, 108 countries, 16M videos, $15M paid |
| Nvidia chip deal via Nscale | Signed September 3, 2025; up to 100,000 chips | $3.5B deal, potentially over $6B |
| Helix 2.5 brain comparison | Trained from zero vs. trained on human videos | 9% success vs. 56% success in 30 homes |
| 30-home robustness test | 13-15 toys, towels, bed-making; strict all-or-nothing scoring | 237 of 420 tries succeeded; best task (bed) ~67% |
| Figure valuation trajectory | 2023 → early 2024 → September 2025 | $0.5B → $2.6B → $39B |
No. Ani Björkström concludes that it is not solved this year or likely next year — a 56% success rate is proof of progress, not a finished product, and Figure itself has said general robotics has not been solved.
Index is a phone app, launched August 2025, that pays ordinary people to film themselves doing housework. It solves a data scarcity problem — there is no "internet of movement" the way there is text for language models — and investors reacted to the resulting performance jump as evidence of a predictable, scalable training pattern rather than a one-off research result.
China accounts for roughly 90% of humanoid robot sales, Tesla is building its own Optimus robot but has not yet made it purchasable, Norway-rooted 1X sells a home robot called Neo for $20,000 or $499 a month, and BMW chose a humanoid robot from Stockholm-listed Hexagon, not Figure, for its first European pilot.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?