AnswersPublished14 sources
Figure Helix 2.5: What Its 30-Home Robot Test Actually Shows
Figure reports that its Index pretrained Helix 2.5 system achieved 56% all or nothing zero shot success on three chores in 30 previously unseen Bay Area homes, versus 9% for a comparable model trained from scratch. The Figure 03 robot tidied living rooms, folded towels, and made beds using a fixed model checkpoint,...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- Figure reports that its Index pretrained Helix 2.5 system achieved 56% all or nothing zero shot success on three chores in 30 previously unseen Bay Area homes, versus 9% for a comparable model trained from scratch.
- The Figure 03 robot tidied living rooms, folded towels, and made beds using a fixed model checkpoint, with no data collection, fine tuning, or adaptation in the test homes or for the objects handled there.
- The evaluation was company run and reportedly used empty, orderly, well lit short term rentals, so its findings should not be read as proof that a robot can safely or reliably operate in lived in homes.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Figure announce about Helix 2.5 and its Figure 03 humanoid robot’s ability to generalize to unfamiliar homes, including the three h. Article summary: Figure announced that Helix 2.5 enabled its Figure 03 humanoid to perform meaningful whole-body household work in homes it had never encountered—not a claim that household robotics is solved. The result is notable eviden. Topic tags: general, general web, education, academic, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, water
Figure’s Helix 2.5 announcement is a notable claim about robot generalization: its Figure 03 humanoid was tested on household chores in homes it had not previously encountered. But the headline number—56% complete-task success—also means the robot did not complete 44% of attempts under the test’s strict scoring rules. 11
4
What Figure tested
Figure says it pretrained Helix 2.5 on Index, its dataset of human behavior, and used the resulting foundation model to create three household behaviors:
- tidying living rooms
- folding towels
- making beds
The company then evaluated those behaviors in 30 previously unseen Bay Area homes. It says no data was collected in any evaluation home, and that there was no site-specific or object-specific fine-tuning or adaptation. One fixed checkpoint was used across the homes. 11
That is the important meaning of “zero-shot” here: the robot was not trained for those particular homes or the particular toys, towels, and bedding it encountered there. It does not mean the three task categories were learned with no task-specific training at all; Figure describes adapting the foundation model into those three behaviors. 11
The reported result: 56% versus 9%
In Figure’s controlled comparison, Index pretraining raised full-task zero-shot success from 9% for an otherwise comparable model trained from scratch to 56% for the pretrained model. 4
11
The company’s grading was all-or-nothing. A tidying run required every scattered toy to reach the basket; towel folding required the towels to be folded and placed; and bed making required the pillows and comforter to be correctly positioned and smoothed. 4
That makes 56% more meaningful than a loose partial-completion score. It also makes the limitation clear: 44% of full attempts failed. The result is evidence that broad pretraining can improve transfer to unfamiliar environments, not evidence that a household humanoid is ready to work reliably without supervision. 4
Why Figure believes the approach can scale
Figure’s broader thesis is that robotics can follow a foundation-model pattern: pretrain on diverse examples of human physical behavior, use a smaller amount of robot data to define a behavior, then deploy it in unfamiliar spaces without rebuilding the dataset for every room. 11
The company reports a scaling relationship in which robot action-prediction loss declined predictably as Index pretraining data increased across an eightfold range. It also said Index was generating roughly 35 minutes of human-experience data per second. 3
A Figure representative separately reported about 90,000 weekly active contributors at filming time, later rising above 100,000. That contributor figure is a company statement rather than independently audited evidence. 28
Figure also announced an initial $3.5 billion compute commitment through its Nscale partnership, with a target of up to 100,000 Nvidia Vera Rubin GPUs and deployment beginning in the second half of 2027. 3
Why the test does not settle household robotics
The result was announced and evaluated by Figure, not established through an independent peer-reviewed household-robotics benchmark. 11
The 30 homes were reportedly short-term rentals that were empty, orderly, and well lit. Real homes introduce people, pets, clutter, variable lighting, interruptions, unusual objects, and safety constraints—conditions that can make manipulation and navigation substantially harder. 4
So the right takeaway is narrower than “robots can now do housework.” Figure has presented a controlled demonstration that pretraining on large-scale human-behavior data may reduce the need to collect robot demonstrations in every new home. Whether that approach can reach the reliability and safety required for ordinary occupied homes remains unresolved. 11
4
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Figure Helix 2.5: What Its 30-Home Robot Test Actually Shows"?
Figure reports that its Index pretrained Helix 2.5 system achieved 56% all or nothing zero shot success on three chores in 30 previously unseen Bay Area homes, versus 9% for a comparable model trained from scratch.
What are the key points to validate first?
Figure reports that its Index pretrained Helix 2.5 system achieved 56% all or nothing zero shot success on three chores in 30 previously unseen Bay Area homes, versus 9% for a comparable model trained from scratch. The Figure 03 robot tidied living rooms, folded towels, and made beds using a fixed model checkpoint, with no data collection, fine tuning, or adaptation in the test homes or for the objects handled there.
What should I do next in practice?
The evaluation was company run and reportedly used empty, orderly, well lit short term rentals, so its findings should not be read as proof that a robot can safely or reliably operate in lived in homes.