Figure AI's Helix 2.5 Cleans 30 Strangers' Homes It Has Never Seen

A single frozen Helix 2.5 checkpoint hit 56% on tidying, towel-folding and bed-making across 30 unseen Bay Area homes, versus 9% without Index pretraining.

Figure AI's Helix 2.5 Cleans 30 Strangers' Homes It Has Never Seen

Figure AI unveiled Helix 2.5, its third-generation vision-language-action neural network for humanoid robots, in a September 17 blog post that reports the first demonstration of zero-shot whole-body autonomy across dozens of previously unseen homes. A single frozen model checkpoint successfully handled living-room tidying, towel folding and bed making in 30 Bay Area homes the robot had never entered, hitting an overall 56% task success rate against 9% for an identically trained baseline that skipped the company's Index pretraining pipeline.

From lab demos to strangers' living rooms

The 30-home run marks the industry's first serious attempt at true generalization for humanoid manipulation. Figure ran 420 attempts split evenly across three household chores: bed making cleared 67%, towel folding 62% and toy tidying 40%, with the same checkpoint deployed unchanged in each home. No home-specific fine-tuning, no operator hand-holding, no fresh data collection on site — the setup a home robot would need to actually ship.

Index dataset does the heavy lifting

The performance jump comes from Figure's Index dataset, which the company says now generates roughly 35 minutes of human-behavior data every second from tens of thousands of weekly contributors. Pretraining Helix 2.5 on Index cuts task-specific data requirements by 50% versus the previous Helix 02 stack, and Figure says robot-action prediction scales predictably with each doubling of Index data — a scaling curve the company is willing to spend $3.5 billion of compute against.

Figure AI Helix 2.5 humanoid robot in a home environment

What Adcock is pitching

CEO Brett Adcock called generalization "the holy grail for robotics: doing work in unseen places," and the 30-home result is the first public data point that Figure's data-first strategy can bend the curve. It also raises the pressure on peers building humanoids for the home tier: NVIDIA's Cosmos world foundation models, Chinese robot-brain startups targeting a 2027 GPT-3 moment and Reward AI's human-data-only OM-1 foundation model are all now measured against Figure's public zero-shot number rather than curated lab demos.

Warehouse-to-home shift underway

Helix 2.5 lands as Figure ramps its BotQ factory, which shipped its 1,000th Figure 03 humanoid in July, and as customer deployments broaden from BMW's Spartanburg logistics floor into more consumer-facing pilots. If the sixfold zero-shot lift the company is claiming holds outside cherry-picked scenes, the road to a genuinely useful home humanoid gets meaningfully shorter.

Reporting based on coverage from Figure AI, The AI Insider, Startup Fortune and Tech Times.

Category: Robotics

Tags: AI Models humanoid robots humanoid AI Physical AI embodied AI VLA Models

Related Articles