Brett Adcock had a problem. His company, Figure AI, had built and deployed a fleet of humanoid robots. The F.02 models logged real hours at a BMW plant in South Carolina. They handled parts, learned tasks, and generated data that fed the next generation. But progress moves fast. The newer F.03 units arrived. Maintaining two fleets no longer made sense.

So the company asked the internet what to do with the old machines. Arnold Schwarzenegger answered. “You should melt them.” The suggestion stuck. What followed was no ordinary scrapyard disposal. Figure trained its retiring robots to walk, leap, and dive autonomously into a 75-ton electric arc furnace filled with molten steel. The event unfolded last week in Imatra, Finland. It mixed practical engineering with cinematic flair. And it underscored a truth about this young industry. Hardware has a short shelf life when software advances rapidly.

The official announcement landed on Figure’s blog on Sept. 30. “We love Figure 02,” the post began. “It is an incredible robot.” The company listed its achievements: first deployment at BMW, birth of the Helix AI system, household chores, logistics work. Yet the math was clear. Expanding the F.03 fleet made older units a burden. Keeping them running demanded resources. So did tearing them apart. Each F.02 contains custom actuators and intellectual property that Figure wanted protected from reverse engineering. Simple recycling wouldn’t suffice. Complete destruction did.

Foundries in the United States and Mexico refused the job. Lithium-ion batteries inside the robots posed too great a risk to their equipment. The team even consulted former MythBusters cast members for ideas. Nothing worked until they located a facility in Imatra willing to accept the challenge. Three massive graphite electrodes power its 75-ton electric arc furnace. The operation offered a tight window: 24 hours, six melts, roughly 20 minutes per cycle before the steel cooled and formed a crust.

Training Machines for Their Own End

Figure didn’t simply push the old robots in. It taught them to exit on their own terms. Engineers set up large airbags at the San Jose headquarters. Robots practiced jumping from a second-floor platform. Stunt performers supplied reference movements. The team built a new AI model in simulation. That model transferred to the real world. The robots had never visited the Finnish foundry. They faced unfamiliar geometry, extreme heat, and powerful electromagnetic fields that fried camera equipment. Still, the machines executed their policies. Some walked forward and stepped off. Others added flair. One performed what looked like a somersault. Cameras caught a final thumbs-up gesture, a clear nod to Terminator 2, before the metal swallowed the frame.

Schwarzenegger showed up for the filming. Dressed in protective gear, he led one robot toward the furnace. Then he stepped back, revealed himself, and delivered the line. “Hasta la vista, Figure 2.” The video, posted widely on X and YouTube, drew millions of views within days. Notebookcheck reported on the marketing stunt Oct. 1, noting how the company turned decommissioning into spectacle. Dexerto covered the same event that day, highlighting the self-termination aspect. The robots didn’t require remote pilots at the critical moment. They acted on trained policies.

But this wasn’t pure theater. The recovered steel shipped back to the United States. Figure is machining it into a limited series of commemorative artifacts now available for purchase. A few F.02 units remain in storage at headquarters. Most of the fleet is gone. The move frees engineering talent to focus on F.04. It also destroys proprietary components that competitors might otherwise study.

The timing reveals broader momentum at Figure. Only weeks earlier, on Sept. 17, the company unveiled Helix 2.5. That neural network achieved a 56% success rate on household tasks across 30 unseen Bay Area homes. No fine-tuning. No prior data from those locations. Bed making hit 67%. Towel folding reached 62%. Tidying proved harder at 40%. A baseline model without Figure’s pretraining on its Index human-behavior dataset managed just 9%. Figure’s own announcement framed the result as evidence that whole-body intelligence can transfer from human data to new scenarios rather than rebuild from scratch each time.

Such gains explain why the company discards hardware so readily. The F.02 robots that once represented the state of the art now serve as raw material. Their actuators and sensors, once prized, become obstacles to iteration. This pattern echoes earlier robotics transitions. Boston Dynamics retired its hydraulic Atlas years ago as it shifted toward electric designs better suited for commercial work. Yet Figure’s approach carries extra symbolism. The company didn’t quietly dismantle the units in a California warehouse. It staged their departure as a public statement about capability and obsolescence.

Challenges remain. Extreme temperatures at the foundry disrupted electronics. Electromagnetic interference knocked out vision systems. The 20-minute windows left little margin for error. Figure acknowledged these conditions in its post. “Despite the extreme heat and electromagnetic fields, the robots ran their AI policies,” it stated, “even as camera equipment and other electronics failed in this environment.” The fact that the jumps succeeded anyway points to the growing strength of the underlying control models. Policies trained in simulation and on airbags held up under real stress.

Investors have taken notice of the pace. Figure’s valuation climbed into the tens of billions on promises of factory and home deployment. A Wall Street Journal investigation last year questioned the gap between claims and current revenue. The company had little income then and few robots in production. Partnerships with BMW provided early validation. The automaker used F.02 units for material handling. Data from those shifts trained later systems. Now that data loop accelerates. Helix models improve. New hardware follows. Old hardware melts.

Recent coverage captured the drama. Wonderful Engineering detailed the logistics on Oct. 1, including the search for a foundry and the precise training steps. The AI Insider noted that the stunt protected trade secrets while generating publicity. On X, reactions mixed awe with unease. One user called the footage “unsettling.” Others praised the engineering. Figure’s own account emphasized authenticity. “In an age of AI-generated footage, we really trained robots to jump autonomously, shipped them to Finland, and had them leap into a vat of molten steel.”

The decommission also highlights practical limits. Humanoid robots cost hundreds of thousands of dollars each. Their components incorporate rare materials and custom designs. When a generation reaches end of life, disposal carries real expense and risk. Figure chose spectacle over silence. The video serves dual purposes. It entertains. It also signals to customers and talent that the company moves decisively. No nostalgia slows development. The next model is always coming.

So what happens to the steel? Some becomes desk ornaments or paperweights etched with the F.02 logo. A tangible reminder that even advanced machines have expiration dates. The rest returns to industrial supply chains. Irony abounds. Robots built to handle steel end up becoming steel. Their final act required new intelligence. The same Helix lineage that lets Figure 03 fold towels in strangers’ homes once guided its predecessors to self-destruct with precision.

Adcock and his team bet that rapid iteration will deliver economic value before reliability catches up to hype. Early factory deployments at BMW produced measurable output: over 1,250 operating hours, 90,000 parts handled, contribution to 30,000 vehicles. Those numbers matter more than any single stunt. Yet the stunt makes the numbers memorable. It transforms accounting decisions into cultural moments. And it leaves an open question for the industry. When your creations can learn to walk into fire, what exactly are you building? Useful tools. Or something that occasionally stares back with a thumbs-up before it disappears.