Figure 03 Review: The Only Humanoid With a Repeat Customer

Almost every humanoid company sells a video. Figure has something better: a customer that ran the robot for ten months and then ordered the next one.

BMW's Spartanburg plant deployed Figure 02 in its body shop, where it supported production of more than 30,000 X3 vehicles over ten months. In June 2026 BMW announced Figure 03 would start at the same plant on complex logistics sequencing. A second-generation deployment at a customer that already ran the first generation is the strongest signal available in this category, and almost nobody else has one.

This review sticks to what BMW and Figure have actually said, and flags separately the numbers that circulate widely and come from neither.

Companion reading: how to choose a humanoid robot, humanoid robot hardware, foundation models and VLAs for robotics, warehouse and logistics robotics, and tactile sensing and e-skin.

Table of contents

1. Key takeaways
2. The number, and what it does not include
3. What Figure 03 actually does at BMW
4. Helix, and why whole-body control matters
5. The hardware changes that matter
6. How it compares
7. What to watch
8. Verdict
9. Frequently asked questions
10. Changelog

The numbers everyone quotes, and where they come from

Two figures circulate constantly in Figure 03 coverage: a cost of around $25 per robot-operating-hour and 99%+ placement accuracy at BMW Spartanburg.

Checked against the primary sources, neither appears in BMW's press release announcing the deployment, and neither is attributed to Figure in it. They trace to third-party analyst write-ups. That does not make them false. It does mean they are not vendor-confirmed operating data, and they should not be quoted as though BMW published them.

If the cost figure is roughly right, two things follow, and they pull in opposite directions.

Why it would matter. No humanoid vendor has published a comparable unit economic. A cost per operating hour is a number a plant manager can hold against a loaded labour cost, which is the conversation the category has been avoiding.

Why it would still not be the whole cost. A per-operating-hour figure typically excludes integration, the cell changes around the robot, supervision, the engineering time to define and maintain the task, downtime while the robot is not operating, and whatever amortisation assumption sits behind it. Comparing it to a wage compares a partial number to a full one.

The honest position: treat both as unverified third-party estimates, and ask Figure directly what the number includes and at what utilisation. The verified evidence for Figure 03 is the deployment history below, and it is stronger than either figure.

What Figure 03 actually does at BMW

Specifics, taken from BMW Group's own release of 25 June 2026.

Figure 03 starts at BMW Group Plant Spartanburg on complex sequencing applications in logistics: delivered components arrive in larger containers and must be ordered so they reach the line correctly.

It follows Figure 02, which BMW describes plainly: over ten months in the Spartanburg body shop, the robot inserted sheet-metal parts for the welding process and supported production of more than 30,000 BMW X3 vehicles. BMW's stated conclusion is that the collaboration "demonstrated that humanoid robots can safely perform precise, repeatable work steps under real production conditions."

Figure's CEO Brett Adcock put the significance in terms worth quoting: "Our 11-month deployment of Figure 02 proved that humanoids are no longer lab experiments." Ulrich Wieland, BMW Manufacturing's VP of Production Control and Logistics, framed the next step as "deploying Figure 03 for a sequencing use case in logistics."

The progression matters more than any single demo. A customer that completed a deployment and then took the successor generation is the strongest evidence available in this category, and it is evidence about the first robot as much as the second.

Note what the work is. Logistics sequencing is structured, repetitive and forgiving of a slow robot. It is exactly where humanoids should start, and it is a long way from the general-purpose machine the category markets. See warehouse and logistics robotics for why this task class falls first.

Helix, and why whole-body control matters

Figure's robots run on Helix, its in-house vision-language-action model. Figure has described successive Helix versions; BMW's deployment release does not name a version, so this section is about the architecture rather than a specific release.

The technically interesting claim is coordinated whole-body control: hands, arms, torso and legs driven together by one policy. Most humanoid stacks are two systems in a trenchcoat, a locomotion controller keeping the robot upright and a manipulation policy driving the arms, with a fragile seam between them. Anyone who has watched a humanoid reach past its balance envelope and abort has seen that seam.

Unifying them is the right architecture and a hard one, because the policy must trade off staying upright against completing the task, continuously. If Helix does this well in production, that is a more meaningful capability than any dexterity demo.

For the model class, see foundation models and VLAs for robotics, and for where the training data comes from, imitation learning.

The hardware changes that matter

The 03 revision reads as a deployment-driven list rather than a spec-sheet chase:

• Soft components across the body, for working near people.
• Wireless charging, which targets availability. A humanoid that docks itself without a person plugging it in is one less reason for the line to stop.
• Tactile sensors and palm cameras in the hands. Palm cameras are the underrated one: seeing into the grasp solves occlusion problems that wrist and head cameras cannot. See tactile sensing and e-skin.
• Speech-to-speech audio, which is mostly an operator-interface feature and matters more on a factory floor than it sounds.

Every one of those is about uptime and working alongside humans rather than about impressing a camera.

How it compares

The contrast with Tesla Optimus is the useful one: vastly more manufacturing ambition at Tesla, and a completed-then-renewed customer deployment at Figure. Different bets, both legible.

What to watch

• Whether the $25/hour figure survives disclosure. Ask what it includes and at what utilisation.
• Whether the German plant pilots convert into deployments like Spartanburg did.
• A customer that is not BMW. One deep customer relationship is a strong signal and a concentration risk.
• Task class expansion beyond logistics sequencing. Structured material movement is the easy end. Assembly is not.

Verdict

Figure 03 is the most credible humanoid deployment in the world right now, and the credibility rests on unusually specific evidence: a named plant, a named task, and a repeat deployment after a prior generation that spent ten months in a body shop and supported more than 30,000 vehicles. BMW said that in its own words, which is rarer in this category than it should be.

Be careful with the numbers that travel alongside it. The cost-per-hour and accuracy figures quoted everywhere are third-party estimates that neither BMW nor Figure published, and a per-operating-hour cost would not be a total cost of ownership even if confirmed. The verified story is better than the unverified one anyway: a customer ran a humanoid for ten months on real production work and then took the next generation. Nobody else in this category can show that.

Frequently asked questions

How much does Figure 03 cost to run? Reported at roughly $25 per robot-operating-hour at BMW Spartanburg in 2026. Read the units carefully: that is per hour of operation, and a figure like this typically excludes integration, cell changes, supervision, engineering time and idle hours.

Is Figure 03 actually deployed or is it a pilot? Deployed, at BMW Group Plant Spartanburg, on complex logistics sequencing. It follows Figure 02 at the same plant, which contributed to assembling more than 30,000 vehicles in 2025. Scope is reported to expand through 2026 and 2027, with separate pilots at BMW's Munich, Regensburg and Leipzig plants.

What is Helix? Figure's in-house vision-language-action model. Figure describes it as doing coordinated whole-body control across hands, arms, torso and legs with a single policy, rather than bolting a manipulation policy onto a separate walking controller.

What is new in Figure 03 versus 02? Deployment-oriented changes: soft components for safety near people, wireless charging for availability, tactile sensors and palm cameras in the hands, and speech-to-speech audio. Palm cameras are the underrated addition, because they see into the grasp where wrist and head cameras cannot.

How accurate is it? A 99%+ placement accuracy figure circulates widely and does not appear in BMW's deployment release. What BMW states, about the predecessor Figure 02 in the body shop, is that the collaboration "demonstrated that humanoid robots can safely perform precise, repeatable work steps under real production conditions."

Can I buy a Figure 03? No. Figure works through direct customer deployments rather than open sales.

Figure 03 or Tesla Optimus? Different stages of the same race. Figure has a completed customer deployment and a renewal for the successor model. Tesla has vastly larger manufacturing ambition and describes its own deployed units as being there to learn rather than to do productive work.