Figure combines humanoid robots, Helix AI and industrial deployment

Figure builds humanoid robots and the AI that makes them act in physical spaces. Figure 03 is its current hardware generation, while Helix is its robot intelligence programme. Industrial agreements provide a commercial route; recent home demonstrations are research evidence. A buyer should keep those two forms of progress distinct when deciding whether a particular task is ready for deployment.

01 / ProductFigure is building both a robot and its learned behaviour

The Figure 03 announcement describes a third-generation humanoid designed around Helix, with redesigned sensing and hands, changes intended for home use, and a manufacturing approach intended to support higher volume. The relevant product idea is that a body built for human spaces can combine movement and manipulation. This is broader than a stationary arm repeating one movement at a fixed workstation.

Helix supplies the learned behaviour. Figure’s September 2026 Helix 2.5 report describes a model pretrained using its Index human-behaviour dataset, then adapted to three tasks and evaluated in previously unseen homes. Figure reports 56% full-task success with Index pretraining versus 9% without it in that experiment. Those are company research results under its stated evaluation conditions, not an independently reproduced reliability figure for a purchased robot.

A third layer is production. BotQ is Figure’s manufacturing programme, including assembly, quality and supporting production software. A model can improve while hardware supply remains constrained; a factory can expand while a task still needs engineering. The purchasing decision therefore depends on behaviour, hardware delivery and support together. None of the three makes the other two automatic.

02 / AudienceIndustrial tasks offer a more concrete starting point than general help

The most concrete public commercial audience is an industrial or logistics organisation with a bounded physical workflow. Figure’s Catalyst Brands agreement describes deployment into a distribution network, starting at a Reno facility. The announcement establishes a commercial relationship and its intended setting. It does not disclose universal pricing, deployment volumes or independently verified returns.

The June 2026 BMW update describes Figure 03 performing a sequencing workflow involving parts and carts at Plant Spartanburg. Figure presents this as a demonstration of its next generation after earlier Figure 02 work. The useful distinction for another manufacturer is whether its own task has a comparable pattern of variation, access and movement. A named factory is evidence of context, not a guarantee that every factory task transfers.

For a household buyer, a robot making a bed or folding towels is understandably compelling. But research progress does not establish an orderable home product, a support network or permission to leave the machine working unsupervised. The reviewed sources do not provide a general consumer checkout or a public retail tariff. Treat domestic demonstrations as a way to understand the research direction, not as a promise that every household chore is supported today.

03 / WorkflowA proposed parts-handling evaluation

Consider a proposed industrial evaluation for sorting non-hazardous sample parts into labelled containers. It would be designed with the vendor and the site’s qualified safety staff, inside an approved test arrangement. This article does not prescribe robot control or safety settings. The aim is to decide whether the task warrants a commercial pilot, and Sequenced has not run this experiment or tested a Figure robot.

Begin with the job as workers actually encounter it. Document the range of part shapes, the variation in container position, the expected completion criteria and the occasions when a person must intervene. The important comparison is against the current process under the same conditions. A robot that succeeds after someone carefully rearranges the objects may be demonstrating a different job from the one the business needs automated.

Use a held-out set of representative sample arrangements. Record full-task completion and the reasons for every incomplete attempt. Separate a mistaken placement, a stalled attempt and a deliberate safe stop: they may all interrupt throughput, but they imply different engineering work. Include the time needed to prepare a new run and recover from an interrupted one. These are suggested acceptance criteria, not measurements from Figure’s deployments.

The BMW sequencing description makes coordination between manipulation and body movement particularly relevant. In the proposed assessment, ask the vendor to identify which parts of the task use established capability and which require new training or integration. This prevents the team from assuming that a similar-looking demonstration already covers the site’s full range of materials and layouts.

End the evaluation with a task-specific decision. It might support a supervised pilot for one repeatable flow while rejecting a broader request to handle every exception. That is a useful result. A precise boundary gives both the supplier and customer a basis for improvement; an undefined ambition to replace a whole role makes cost and performance much harder to assess.

04 / PricingCommercial partnerships do not supply a public unit price

The reviewed primary material shows agreements and deployments, not a published catalogue price for Figure 03. The Catalyst announcement is a commercial agreement; the BMW update describes a specific application. Neither establishes a standard hourly rate, lease payment or purchase price available to all buyers on 22 September 2026.

A useful proposal would separate the robot, task adaptation, site integration, maintenance and any ongoing software or service obligations. Those are suggested cost categories, not claims about Figure’s billing structure. Ask what happens when the task changes, when a component needs replacement and when the robot cannot complete a scheduled shift. A low hypothetical hardware price is not a meaningful budget if those responsibilities remain undefined.

05 / DistinctionsGeneralisation and manufacturing are different advantages

Figure’s research matters because robots encounter combinations of objects, layouts and actions that are difficult to enumerate manually. The Helix 2.5 report makes its evaluation boundary unusually useful to readers: zero-shot refers to the held-out homes and objects, while tasks were specified using data collected elsewhere. It does not mean the robot learned every possible household job without task-specific work.

The manufacturing approach addresses a different problem. BotQ’s announcement describes redesigning components and production processes for scale and bringing core assembly in-house. Its stated capacity ambitions should not be read as a count of robots shipped. For a buyer, the relevant evidence is the delivery and support commitment in its own programme, not the size of a long-term manufacturing aspiration.

Compare the ownership model with the NVIDIA blueprint, which concerns enabling AI software and infrastructure rather than purchasing this integrated humanoid system. The Hugging Face blueprint is useful when the job is researching and distributing model resources. Figure’s public research is informative, but it does not establish that Helix weights, training data or robot interfaces are offered through a comparable open developer route.

06 / QuestionsWhat must be resolved beyond the demonstration

The central uncertainty is the gap between a successful example and sustained useful work. A full evaluation needs the distribution of outcomes: how often work completes, how often people assist, and which conditions explain the difference. A research success rate under one rubric cannot be multiplied by operating hours to produce a credible labour saving. The business process may tolerate some interruptions and be highly sensitive to others.

Another question is whether a task can remain stable enough to justify deployment. A warehouse that changes packaging weekly may face a different adaptation burden from one handling a consistent parts family. Ask how new objects and layout changes are introduced, who approves them and what evidence accompanies an updated behaviour. Public material does not answer those contract-specific questions.

Hardware service also deserves separate attention. Hands, batteries and other physical components create maintenance obligations that software demonstrations cannot reveal. Request the planned support geography, spare-parts arrangement and downtime responsibilities. Avoid assuming a manufacturing programme automatically supplies local field service.

Finally, determine which commercial opportunities are currently open to the reader. The public evidence supports industrial engagement and active research, but not a universal promise of consumer availability. Figure’s latest work can justify a serious discussion while still leaving the right decision as “monitor progress” for a household or small business that needs a ready-made general assistant.

07 / DecisionSelect a task with a measurable boundary

Figure is a useful company to follow because it is advancing learned robot behaviour alongside physical product design and manufacturing. For an industrial buyer, the strongest next step is a task-level evaluation with explicit completion and intervention criteria. For a consumer, the right question remains whether there is a supported commercial offer for the specific job, rather than whether the latest research video is impressive.