When a robotics company outgrows a whole generation of machines, the usual story is scrap yards and screwdrivers. Figure chose something louder. It taught its older humanoid robots — the Figure 02, also called F.02 — to leap into molten steel, then turned what came out into keepsakes you can buy.
Why the old fleet had to go
Figure is an AI robotics company building humanoid robots — machines shaped roughly like people, with arms and legs that can work in human spaces. Its Figure 02 robots marked a long list of firsts for the firm: an early factory deployment at BMW, the start of Helix (the company’s AI software stack), household chores, and logistics work.
Those milestones mattered. So did the next generation. As the newer F.03 fleet grew, keeping the older F.02 machines running stopped making sense. On September 30, 2026, Figure said so plainly in its own post: the company was decommissioning the F.02 fleet.
Ordinary scrap would not do. The robots were packed with custom actuators — the motors and joints that make limbs move — and other proprietary hardware. Figure needed total destruction so that intellectual property (the hard-won designs inside the machines) would not leak out through a recycling chain. Taking every robot apart by hand would have tied up technical staff long enough to delay the next model, F.04. So the company looked for a faster, final answer.
The internet — and Arnold — said melt them
Figure asked the public what to do. The loudest reply came from Arnold Schwarzenegger, who told them to melt the robots. The company reached out. When he said he wanted to be involved, the plan moved from joke to project.
That is a strange origin story for an industrial disposal job. It also fits the moment. Humanoid robots are still rare enough that retiring a whole generation feels like news, not routine warehouse cleanup. Melting them — on camera — made the IP problem theatrical and, Figure hoped, airtight.
No foundry in the U.S. or Mexico would take the jump
Here is the everyday obstacle: lithium-ion batteries. Those are the rechargeable packs that power phones, laptops, and, in this case, walking robots. Foundries in the United States and Mexico refused to let a robot carrying those batteries jump into their furnaces. Expensive equipment and battery fire risk do not mix.
Figure even asked alumni of the TV show MythBusters to help hunt for a willing site. The only foundry that said yes was in Imatra, Finland.
Training the jump in San Jose
Before anyone flew to Finland, the team trained at Figure’s campus in San Jose, California. They set up large airbags and taught robots to jump off the second floor of a building. Stunt artists’ movements became a reference — a human blueprint for how a body launches and lands.
Then came simulation. The company trained a new AI model in a computer world so the robots could aim a jump into a bucket of molten steel at a foundry the machines had never visited. In plain terms: the software practiced the leap thousands of times in a fake foundry so the real leap could land in the right place on the first try.
That detail matters more than the stunt footage. Walking into a room you have never seen, then hitting a small target under time pressure, is exactly the kind of skill humanoid robots are supposed to learn. Figure used retirement as a stress test for that skill.
Twenty-minute windows in a 75-ton furnace
At the Finnish foundry, three massive graphite electrodes powered a 75-ton electric arc furnace — a vessel that melts steel scrap with huge electric arcs rather than a traditional flame. The team had 24 hours and six melts. Each melt gave them about a 20-minute window before the molten steel cooled and formed a hard crust on top.
Heat and electromagnetic fields (the invisible disturbances strong electric currents create) were brutal on ordinary gear. Camera equipment and other electronics failed. The robots, Figure says, kept running their AI policies — the learned rules that tell them how to move — and completed the jumps anyway.
Gizmodo’s reporting added a celebrity beat: Schwarzenegger appeared in the film with a thumbs-up and a line that echoed his old movie catchphrase, aimed at the retiring machines.
Bars, artifacts, and what is left
After the melt, the resulting metal bars shipped back to the United States. Figure is machining them into a limited run of commemorative artifacts — small objects meant to preserve a piece of the company’s early hardware history. Gizmodo reported preorders in roughly the $500–$1,900 range, with the pricier pieces sold out. That price band comes from reporting, not from Figure’s own blog post.
Most of the F.02 fleet is gone. A few robots remain in storage at headquarters. Figure insists the footage is real, not AI-generated: the company trained the jump, shipped the machines to Finland, and had them leap into molten steel. In an era when fake video is cheap, that insistence is part of the story.
A retirement that doubles as a demo
Plenty of tech companies retire hardware quietly. Figure turned the goodbye into a capability demo. Teaching a humanoid to jump accurately into a moving industrial target is not the same as walking a warehouse aisle, but it uses the same ingredients: perception, balance, timing, and a policy trained mostly in simulation.
That is why the company keeps stressing that the robots ran their AI policies at the foundry even while cameras died. The stunt sells the melt. The reliability claim sells the product line that comes next. Viewers who only remember the Terminator wink still leave with one sticky fact: these machines were not puppeteered frame by frame for a commercial. Figure says they jumped on their own, after training.
None of that answers every hard question about humanoids — cost, reliability over years, or what happens when a battery-packed machine fails in a real factory. It does answer a narrower one: when Figure needed the F.02 generation gone, it found a way that protected its designs, produced a limited set of metal artifacts, and put the next fleet under a brighter spotlight.
What this episode actually shows
Strip away the Terminator jokes and you still have three practical points.
First, custom hardware creates a disposal problem that scrap yards are not built for. If your competitive edge lives in actuators and boards, you cannot casually recycle the next generation into someone else’s supply chain.
Second, batteries change what industry will allow. A foundry that melts steel all day still drew a hard line at lithium-ion packs diving into the melt. Only one site worldwide said yes.
Third, simulation-to-real transfer — teaching a skill in software, then running it in a place the robot has never seen — is no longer a lab demo. Figure used it for a one-shot industrial stunt under a twenty-minute clock.
You do not need to love the spectacle to notice the engineering. A company that once celebrated F.02 as a breakthrough chose to destroy most of that fleet so the next generations could move faster. The jump into molten steel was protection theater with a furnace, and also a public exam for an AI that had to stick the landing.
For the company’s own telling and film, see Figure’s F.02 decommission post.
Further reading
Artificial Intelligence: A Modern Approach — The classic Russell and Norvig textbook on AI ideas, including robotics and decision-making, for readers who want the foundations behind humanoid policies.
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow — A practical guide to the learning techniques that power modern robot policies, from basics through deep networks.