A humanoid robot walks onto a diving board and throws itself into a vat of molten steel. Standing nearby in the footage: Arnold Schwarzenegger — the same way he terminated himself in Terminator 2.
This isn't a movie. On September 30, Figure published this decommission notice under a two-word title: F.02 Decommission.
Why should you care about a vat of steel? Because the official material around it carries numbers you never hear at a launch event: no foundry in the United States or Mexico would take the job, and taking the robots apart one by one would slow down the next generation. In Figure's April production update, line takt time rose 24-fold in under 120 days. A company spending real money to melt down its own previous generation and then selling you the leftover metal as a souvenir says more than any humanoid launch event this year about how fast this industry is moving — and how badly it needs you to believe it really is moving.
This "launch event" has no new product — just a furnace and a shopping link
Conventionally, news like this comes with a new robot. But the official notice has no new product and no spec sheet, only three things: a video, an explanation, and a store link for limited-edition artifacts.
The notice lays out three things, and only two of them are reasons for retiring the fleet.
First, once the F.03 fleet was scaling, keeping the F.02 fleet running no longer made sense. Second, disassembling the robots one by one would tie up technicians long enough to delay F.04.
The third item is about disposal, not about why. Figure says it needed to destroy the robots in a way that protected its proprietary hardware, and no foundry in the U.S. or Mexico was willing to drop robots carrying lithium batteries into its expensive equipment. Figure even asked veterans of MythBusters for help finding a site; in the end, only one foundry, in Imatra, Finland, said yes.
As for why "melted," Figure's notice says the company asked the internet: the answer came from Schwarzenegger himself, who said to melt them. The company then reached out and invited him to take part. That part is Figure's own story.
Put plainly, this is a content drop. But it carries more information than a typical one, precisely because making "we're not just putting on a show" believable requires laying out cost, capacity and timelines.
The previous generation had to go — not because it broke
The line that matters in the notice is this: taking the robots apart one by one would delay F.04. What sealed F.02's fate wasn't a breakdown, and it wasn't a lack of intelligence; it's that taking one apart costs more than throwing it away.
That's different from how a carmaker handles last year's model. An old car can be discounted to clear inventory, or sold on to a rental fleet. A robot can't: its body is full of structures the company never meant to make public.
And F.04 is long past the paper stage. According to Gasgoo reporting, Figure founder Brett Adcock said on X on May 13 that F.04's design was fully locked and parts delivery had begun, calling it "the biggest leap between generations."
Meanwhile the line keeps speeding up. In its April production update, Figure said the BotQ factory had delivered more than 350 F.03 units, with takt time going from one a day to one an hour — a 24-fold gain in under 120 days. The plant runs more than 150 connected workstations and more than 50 in-process inspection points, with first-pass yield above 80% at the end of the line, 99.3% on the battery line, and more than 9,000 actuators produced.
| F.02 (previous generation) | F.03 (current) | |
|---|---|---|
| Manufacturing | Figure says mainly CNC machining | Figure says heavy use of die-casting, injection molding, stamping |
| Build cost | Baseline | Figure says only "dramatically less," no figure; industry press estimates about 90% lower |
| Capacity | Not disclosed | Figure says takt went from 1/day to 1/hour |
| Milestones | First humanoid in a BMW plant, plus household and logistics work | Per industry press, unit 1,000 came off the line in July |
| Status | Figure says most no longer exist; a few remain at HQ | In production, still scaling |
Two entries need their sourcing spelled out. The manufacturing row comes from Figure's Figure 03 page from October 2025: the previous generation was mainly CNC-machined, while the new one leans heavily on die-casting, injection molding and stamping. On cost, Figure's own wording is only that "each Figure 03 unit now costs dramatically less to build" — no percentage. The roughly 90% figure is an industry-press estimate.
Put those together and the conclusion is blunt: the new generation is cheaper, faster, and still scaling. Old robots sitting in a warehouse generate no data and no revenue — only cost.
It doubled as a stress test, but it's not a performance result
From an engineering standpoint, there is something real in this furnace run. Figure says the training started at its San Jose campus: the team set up large airbags and taught the robots to jump from the second floor of a building, using stunt performers' motions as a reference, and trained a new model in simulation so the robots could land in the steel accurately at a foundry they had never visited.
On site: a 75-ton electric arc furnace, three huge graphite electrodes, six melts in 24 hours, and only a 20-minute window per melt — after that the steel cools and crusts over. Figure says that under extreme heat and strong electromagnetic fields, the robots' AI policy ran normally, and it was the camera gear and other electronics that failed first.
This is the closest thing to technical information in the whole affair: a robustness check in a real environment. Extreme heat, strong electromagnetic fields, and the robot has to complete the motion on its own.
But it isn't a performance evaluation. No third party was present, there was no repeated testing, and there are no comparable specs. It shows the robot didn't crash in that environment; it says nothing about how well F.02 could work in a factory, and it shouldn't be used to rank it against any robot actually on the job.
One dispute is worth stating clearly. Early in the video, a robot lowers itself down a chain. Industry outlet Futurism judged that this sequence "very likely" involved human assistance, and Figure hasn't explained it frame by frame. The only thing the company emphasizes is that the robots jumped into the steel on their own and that the footage is not AI-generated.
That second point deserves its own line. In 2026, a robotics company felt it had to state in a notice that "this is not an AI-generated clip." That alone tells you where the industry stands: audiences now assume any footage might be synthetic, so proving you really did the thing has become something you have to spend serious money on.
From scrap steel to souvenirs: turning destruction into product
Figure says the metal bars from the melt were shipped back to the U.S. and made into a small run of limited-edition artifacts, with a purchase link right in the notice.
You can't buy the robot, but you can buy part of it. The prices are public: two tiers in Figure's store, a 2/3-scale large piece at $1,900 and a 1/3-scale small piece at $500. The small one is available for pre-order with estimated shipping in 6–8 weeks; the large one's page shows it can no longer be purchased and, according to a Paper report carried by Beijing Daily, it sold out. Both product descriptions say each piece comes from metal recovered in the 75-ton arc furnace.
What Figure hasn't disclosed is two other numbers: how many robots were melted, and how many pieces were made in each tier.
The point of this accounting is brand. It adds no capacity and generates no data, but it converts assets that would otherwise have vanished into a sense of belonging. Hardware companies increasingly look like content companies: a launch is content, and a decommission is content — the difference being that a decommission gets shared more.
In the same race, others are comparing who ships more
Put this show back into the global race and another set of numbers cools it down fast. According to "Automotive Semiconductor" on EET China reporting, UBTech announced its 15,000th robot off the line in June 2026, and Unitree disclosed humanoid-only shipments above 5,500 units in 2025. In the same article, Figure's milestone is 1,000.
These numbers aren't measured the same way. Some are "off the line," others are "shipped"; some are cumulative, some cover a single year. When industry outlet Korthos Robotics recorded Figure's milestone, it noted specifically that 1,000 means "built," not deployed at customer sites. They're meaningful side by side, but they shouldn't be read as a simple ranking.
That's the other side of the furnace stunt: attention and volume are two different things. The video proves Figure tells a good story; 1,000 units prove it's starting to build. You need both, but only one turns into revenue.
Next time you see news like this, ask three questions
A decommission notice often carries more information than a launch notice. A launch shows only what it wants you to see; a decommission has to explain why the thing isn't being kept — and "why not keep it" automatically drags in cost, capacity and roadmap.
Next time you run into a similarly showy item, ask three questions first.
Was this number "built" or "sold"? Off the line, shipped, delivered and working on site are four different things, and units drop out at every step.
How much of what's on screen is the robot deciding, and how much is the set and the crew? With no third party present, treat it as promotional material, not a test report. This time even the company had to declare the footage wasn't AI-generated — a reminder of how far "looks real" still is from "can be verified."
What hasn't the company said? This time it's how many robots were melted and how many pieces were made per tier. What goes unsaid is usually what it doesn't want you to calculate.
If you follow humanoids, the F.02 meltdown is a more direct signal than any spec sheet: this generation's competition is no longer "can you build one" but "can you clear the last generation out of the warehouse and turn the next one into a production line." Whoever still has the previous generation stacked in a warehouse is paying tuition with its cash flow.
Sources
- Figure official notice: F.02 Decommission (September 30, 2026)
- Figure official notice: Ramping Figure 03 Production (April 29, 2026)
- Figure official store: F.02 Artifact (pricing and availability)
- Further reading: Futurism, Horde of Obsolete Humanoid Robots Made to Hurl Themselves Into Vat of Liquid Steel
- Further reading: EET China, "Automotive Semiconductor" column, Unit 1,000 rolls off the line: Figure 03 humanoid hits a production milestone
- Further reading: Gasgoo, Figure founder: F.04 is entering parts delivery
- Further reading: Korthos Robotics, Figure Reaches 1,000 F.03 Humanoids as BotQ Production Accelerates
- Further reading: The Paper (republished by Beijing Daily, October 1, 2026), original report
- Further reading: Figure official video, F.02 Decommission
