Figure AI is a US humanoid robotics company founded in 2022 and led by CEO Brett Adcock. It develops general-purpose bipedal humanoid robots paired with proprietary Vision-Language-Action (VLA) models, of which its flagship is Helix.
| Field | Value |
|---|---|
| Company type | Robotics (humanoid) |
| Founded | 2022 |
| HQ country | US |
| CEO | Brett Adcock |
| Flagship product | Figure humanoid robots (third generation: Figure 03) |
| Flagship model | Helix (Vision-Language-Action model) |
Models
Figure introduced Helix, described as "A Vision-Language-Action Model for Generalist Humanoid Control," on February 20, 2025. Helix unifies perception, language understanding, and motor action in a single model, analogous to how GPT-4V unifies modalities but extended to physical actuation. It is the first publicly announced VLA model specifically targeting generalist humanoid control, a direction distinct from domain-specific robotics controllers.
Figure introduced Helix 2.5 on September 17, 2026. Unlike its predecessor Helix 02, which started from a pretrained vision-language model and learned behaviors from data collected where the robots would operate, Helix 2.5 was pretrained from random initialization entirely on the company's Index dataset of human behavior, then adapted to three whole-body behaviors: tidying living rooms, folding towels and making beds. Figure evaluated it in 30 Bay Area homes where no data had been collected, with unseen objects and no partial credit, and reports that Index pretraining raised zero-shot success from 9% for an otherwise identical policy trained from scratch to 56%. It also reports that Helix 2.5 matched a Helix 02 policy's success rate with half as much task-specific data, and that downstream action-prediction loss fell predictably with each doubling of Index data across an 8× range. The company states the result does not mean general humanoid robotics is solved; the figures are its own and are not independently verified (Source: figure.ai). See Helix (Figure AI Vision-Language-Action model).
Hardware
Figure's robots have progressed through three generations, with Figure 03 the third-generation platform in deployment by mid-2026. The company positions Figure 03 both for industrial work and, prospectively, for the home, running on the Helix model for perception and control (Source: eletric-vehicles.com).
Partnerships and commercial status
Figure has announced a manufacturing partnership with BMW and run logistics pilots. At a technology conference in June 2025, Adcock skipped a live demo and sidestepped questions about the BMW partnership, which prompted skepticism about the real-world operability of Figure's platform at manufacturing scale (Source: https://techcrunch.com/2025/06/06/figure-ai-ceo-skips-live-demo-sidesteps-bmw-deal-questions/). The skipped-demo episode has been cited as evidence of over-promising, and commercial traction remains under scrutiny.
Through 2026 the company reported moving from pilots toward standing deployments. In early July 2026 Figure said it had deployed its third-generation humanoid to perform logistics work at a BMW Group plant in the United States, extending the earlier BMW relationship from pilot testing to production-line tasks (Source: eletric-vehicles.com). Separately, Figure robots were reported in May 2026 to be taking on work tied to retail and apparel operations associated with JCPenney, Aéropostale and Brooks Brothers (Source: forbes.com).
Data collection
Figure came out of stealth on August 25, 2026 with Index, a crowdsourced platform for collecting physical-task video to train the Helix stack, launched on Google Play and the App Store after four months of stealth operation. The company reports 264,000 downloads across 108 countries, more than 44,000 weekly active users, more than 16 million uploaded videos, and processing of 30 minutes of video every second, which it characterizes as 4.9 years of human work uploaded each day. Per 1,000 hours collected it counts 373 unique tasks, 1,146 unique manipulated objects and 116 unique environments. Figure says it has paid $15 million to contributors, whom it calls Creators, and has committed to spending more than $1 billion on data and compute over the following twelve months. Submissions pass a five-stage pipeline of automated filtering for technical, visual and semantic quality; human fraud review auditing samples at the user level; embedding-based deduplication discarding segments above a similarity threshold with accepted data; rebalancing by task quota and embedding cluster; and hierarchical text captioning of every episode. The Creator model runs in two directions: a contributor can record tasks in their own home or workplace, or book a Creator through the app to come to their home or business and perform tasks while recording. The diversity counts are stated per 1,000 hours collected rather than as dataset totals, and the post's summary bullets give "100+ countries" where its body gives 108 (Introducing Index: Building The World's Largest and Most Diverse Physical Dataset (Figure AI, August 2026); glitchwire.com).
The company states it first tried buying data and found that vendors could not meet its throughput, diversity or quality requirements, that it is "now on a path to 100x," and that Index "is laying the groundwork for ordering robots as a service." All figures are Figure's own and are not independently verified. The Creator programme is a paid data-collection workforce; see Data-labeling and data-collection labor.
By the Helix 2.5 announcement of September 17, 2026, Figure reported that Index was generating roughly 35 minutes of new human experience every second, against the 30 minutes of video per second given at launch, and that it had committed $3.5 billion of compute to training Helix, a commitment the post links to Figure's strategic partnership announcement with Nscale (Source: figure.ai). Crusoe named Figure among the customers of its Crusoe Cloud service in its September 17, 2026 funding announcement (Source: crusoe.ai).
The approach contrasts with the in-house multi-source strategy Skild AI describes for S1, which rates teleoperation, hand-held collection, egocentric video and simulation against hardware proximity, diversity and scalability and scales all four internally (Introducing S1: In-Context Learning for Robotics).
Public profile
In March 2026 the White House hosted a Figure humanoid robot, with First Lady Melania Trump appearing alongside it in what was reported as the first humanoid-robot guest at the White House — an indication of the visibility humanoid robotics had reached in US technology-policy discussion (Source: cnbc.com).
Industry context
Figure is described as one of four frontrunners in the humanoid-robotics sector, alongside Physical Intelligence, Tesla Optimus, and Clone Robotics.
Relationships
- deploys-in: industrial/manufacturing (BMW partnership and 2026 US-plant logistics deployment), plus retail/apparel and logistics pilots.
- related: The Industrial Explosion (Davidson, Hadshar) (the Davidson/Hadshar theory frames humanoid robotics as the industrial-explosion substrate), Physical Intelligence, Clone Robotics, Amazon (Amazon humanoid robot testing), Helix (Figure AI Vision-Language-Action model), AI Robotics.
Sources
- Figure / Helix: A Vision-Language-Action Model for Generalist Humanoid Control (2025-02-20)
- TechCrunch / Figure AI CEO skips live demo, sidesteps BMW deal questions (2025-06-06)
- CNBC / Meet Figure AI: the company behind the humanoid robot hosted by Melania Trump (2026-03-26)
- Forbes / Figure Humanoid Robots Get Jobs With JCPenney, Aéropostale, Brooks Brothers (2026-05-26)
- Electric-Vehicles.com / BMW Deploys Figure's New Humanoid for Logistics at US Plant (2026-07-01)
- Figure / Helix 2.5: Zero-Shot 30-Home Generalization (2026-09-17)
- Crusoe / Crusoe Raises $3.9 Billion Series F (2026-09-17)