The humanoid robot entering the factory floor.

A human shape, an industrial question

The interesting thing about a humanoid in a car factory is not that it resembles a person. It is that the factory already has robots. Figure's experiment asks whether a machine with legs, hands and learned behaviour can take on work between established automation systems, rather than requiring a production line to be rebuilt around another fixed machine.

At BMW's Spartanburg plant in South Carolina, that question initially took a specific form: moving sheet-metal parts into fixtures so conventional industrial robots could carry out the next operation. BMW described Figure 02 trials in a real production environment in 2024. The task was modest compared with the promise of a general-purpose humanoid, but useful precisely because it could be measured against an existing process.

Figure is building both the humanoid hardware and the artificial intelligence intended to control it. Its ambitions reach from commercial workplaces to homes. The distance between those ambitions and a dependable product is the central story: a machine must do more than demonstrate a skill once; it must keep earning its place in a workflow.

The founder's wager on physical work

Brett Adcock founded Figure in 2022 after previously founding Archer Aviation and Vettery, according to the company's launch announcement. When Figure emerged publicly in March 2023, it presented Figure 01 as the first machine in a programme aimed at labour shortages and work it described as unsafe or undesirable. Those were the company's reasons for building, not evidence that its robots had already solved either problem.

The original announcement said the team had completed a full-scale humanoid build and would begin testing in the following months. That chronology matters: Figure 01 was an engineering starting point, not an established commercial workforce. Manufacturing, logistics, warehousing and retail appeared as intended applications, giving the company a broad destination but leaving individual jobs to be proved separately.

The headquarters city is San Jose, California: Figure's own careers page explicitly identifies it as such. The company's early announcements carried a Sunnyvale dateline, which can make older profiles misleading. More important than the address is the scope of the undertaking housed there: Figure is attempting to develop a physical product and its learned control system together, rather than treating intelligence as an attachment to finished machinery.

Three bodies, a moving design brief

Figure 01 established the original humanoid proposition. Figure 02 became the version associated with BMW's early factory work. BMW's 2024 account described improvements over its predecessor in computing, battery capability and communication, alongside hands intended for precise manipulation. These changes addressed practical constraints: seeing the work, reaching it, holding a component and staying powered through useful operations.

Figure introduced its third generation, Figure 03, in October 2025. The company described a ground-up redesign around Helix, its vision-language-action AI, with revised cameras and hands. It also described palm cameras, tactile sensing, soft coverings and wireless charging. These are company-described design features, not independent proof that the machine is ready for unsupervised household use.

The shift towards the home broadens the difficulty. A production fixture offers a comparatively bounded interaction; domestic spaces contain changing objects, furniture and people. Figure says Figure 03 was also designed for high-volume manufacture. That makes the generation change more than a new exterior: it combines a broader operating environment with a requirement to build many consistent copies of the same machine.

Helix is the other half of the robot

Helix, announced in February 2025, is Figure's attempt to connect what a robot sees and understands with the movements it performs. The company described a vision-language-action model controlling the upper body, including wrists and individual fingers, and showed two robots putting away groceries. Its claim was that learned behaviour could extend across unfamiliar objects without a separate, manually programmed routine for every item.

That is an important distinction from simply replaying a motion. The physical arrangement can change while the instruction stays the same. Figure's demonstrations make the research direction tangible, but a demonstration cannot establish reliability across all the objects and circumstances suggested by a general instruction. The scope of a model's apparent understanding and the scope of its dependable behaviour are different questions.

In January 2026, Figure announced Helix 02, extending its stated control approach to walking, manipulation and balance together. It presented a four-minute dishwasher task as autonomous, with no resets or human intervention. This was a company-reported demonstration, not a household service record. Nevertheless, it illustrates why software and body must develop together: reaching for a dish can require coordinated changes in stance, vision and grip.

What BMW establishes—and what it does not

BMW's own account confirms that Figure 02 placed sheet-metal parts into special fixtures during testing. Figure's November 2025 retrospective went further, reporting more than 90,000 parts loaded, over 1,250 hours of runtime and a contribution to more than 30,000 X3 vehicles. Those totals are Figure's reported deployment results. They do not mean a humanoid assembled complete cars or displaced the rest of the production equipment.

The timeline deserves care. In April 2025, Fortune reported, citing a BMW spokesperson, that the work had until March taken place during off-hours and had subsequently moved into live production with a single robot doing a limited task. That contemporaneous account should not be erased by later cumulative numbers; nor should an early, restricted deployment be treated as proof that subsequent progress was impossible.

Figure's retrospective identified cycle time, placement accuracy and human interventions as critical measures. It listed an 84-second total-cycle requirement and a placement-success target above 99 percent per shift. These were requirements and targets, not interchangeable with independently verified achieved averages. The distinction is essential when judging whether a useful experiment has become dependable production equipment.

The factory beyond the demonstration

Manufacturing the robots is a second industrial challenge. In April 2026, Figure said its BotQ facility had delivered more than 350 Figure 03 robots and demonstrated a production cycle of one robot per hour. The company described supplier qualification, inspection points and dedicated module lines. These are reported manufacturing milestones; they do not establish how many robots were working for paying customers or how consistently the stated production rate was sustained.

In June 2026, Figure reported that Figure 03 had arrived at BMW for a sequencing workflow. Its description included selecting and placing parts as well as pulling a wheeled cart, with Helix 02 coordinating body movement and manipulation. The announcement presented a demonstration of a new application, rather than publishing a comparable long-duration operating record for the newer robot.

The next test is therefore not merely another video of dexterity. It is whether different units can perform the same useful work with predictable servicing and little interruption. Figure's accounts show why this matters: the BMW task demanded speed and placement precision together, while its manufacturing report stressed yield and quality checks. A learned skill and a repeatable industrial product have to meet in the same machine.

Capital is not the same as unit economics

Reuters reported in September 2025 that Figure said it had secured more than $1 billion in committed capital in a Series C financing at a $39 billion post-money valuation. Parkway Venture Capital headed the round, with participants including Nvidia and Intel Capital. The wording matters: committed capital and a financing valuation describe investor backing, not revenue, profitability or the price customers will pay for useful robot hours.

That backing gives context to the breadth of Figure's effort: robot design, model development and manufacturing all require resources. But the sources reviewed here do not establish a public, independently audited cost per productive hour for the BMW work. Without operating costs, maintenance demands and intervention rates over time, a reader cannot turn a deployment milestone into a credible payback calculation.

The business stands at the intersection of visible technical progress and still-open commercial questions. Figure has identified real industrial tasks and reported production activity. Its larger proposition—that a shared humanoid platform can become useful across many workplaces and homes—remains broader than the evidence from any one installation.

Why Global58 noticed them

Global58 noticed Figure because it makes an unusually concrete question visible: can general-purpose AI become dependable physical labour in environments designed around people? The interest is not in the human silhouette alone. It is in the combination of learned control, hands, mobility and the unglamorous demands of a production process. This editorial classification is not a ranking or endorsement.

What matters next is comparable evidence: longer operating records for Figure 03, clearly stated task boundaries, intervention frequency and customer-confirmed results. Household demonstrations should be assessed separately from industrial deployments, and manufacturing output separately from productive customer installations. Figure is worth following because these distinctions can now be tested against physical work—not because a convincing glimpse of the future guarantees that the whole future has arrived.