What Helix is
Helix is a robot control model family from Figure for humanoid perception and whole-body action control. On September 17, 2026, Figure announced Helix 2.5, calling it the most advanced neural network the company has built. It covers whole-body behaviors including locomotion, rigid and deformable manipulation, bimanual coordination, and active perception, with tidying living rooms, folding towels, and making beds as its validation tasks.
Capabilities and boundaries
The core result in the official announcement is zero-shot generalization: one fixed checkpoint completed the tasks across 30 real homes with no data collection, fine-tuning, or adaptation in those environments or on those objects. Figure describes this as the first demonstration of zero-shot whole-body generalization at this scale on a humanoid, and quantifies the pretraining effect: pretraining on its Index data engine raised zero-shot success from 9% (a from-scratch policy) to 56%. The company also states that behavior specification became 2x cheaper while its scope expanded 30x.
In training terms, Helix 2.5 differs from the earlier Helix 02: Figure states that Helix 02 started from a pretrained vision-language model, whereas Helix 2.5 was pretrained from random initialization entirely on Index.
These results come from Figure’s own evaluation (30 Bay Area homes, a single checkpoint, “no partial credit” scoring) and are vendor-reported. Official materials describe Helix in the context of Figure’s own humanoid robots; the announcement offers no licensing, open download, or API for outside parties.
Discussion in the show
In Weekly #003’s chapter “具身智能与零样本任务” (embodied intelligence and zero-shot tasks), Guizang relays the Helix 2.5 release in the Chinese transcript: the main gain is zero-shot task success, lifted from 9% to 56% for tidying rooms, folding towels, and making beds, and he argues that progress in embodied intelligence — “the brain part” — remains slow. Orange adds that the demo video shows the robot picking socks off the floor, which he found slightly eerie but interesting. At the end of the previous chapter, Guizang refers to the release as coming from “the robot company that previously worked with OpenAI,” and Orange confirms it is Figure. The episode 003 chapter has the full context; an English transcript is not available.
Frequently asked questions
What is Helix?
Figure’s robot control model family for humanoid perception and whole-body action control. Helix 2.5 was announced on September 17, 2026.
What does Helix 2.5 improve over earlier versions?
Per the official announcement: zero-shot whole-body autonomy across 30 real homes; Index pretraining raised zero-shot success from 9% to 56%; behavior specification became 2x cheaper while its scope expanded 30x.
How do Helix 02 and Helix 2.5 differ?
Per Figure: Helix 02 started from a pretrained vision-language model, while Helix 2.5 was pretrained from random initialization entirely on the Index data engine and matched a Helix 02 behavior with half the adaptation data.
Can other companies or researchers use Helix?
The announcement offers no licensing, open download, or API; the published results all concern Figure’s own humanoid robots.
Are the reported numbers independently verified?
The success rates and generalization claims were evaluated and published by Figure itself (30 Bay Area homes, a single checkpoint, no partial credit); the official page lists no independent replication.