Figure AI Robot Decommissioning Ends in Molten Steel, but the Hard Part Comes Next

Figure AI robot decommissioning took an extraordinary turn when retired F.02 humanoids jumped into a 75-ton furnace filled with molten steel. The company says it destroyed most of the fleet because maintaining or dismantling those machines would divert engineers from newer models.

The robots traveled to a foundry in Imatra, Finland, for a sendoff modeled after the ending of Terminator 2. Arnold Schwarzenegger suggested melting them after Figure CEO Brett Adcock asked online for retirement ideas.

The resulting film is memorable, but the underlying decision matters more than its cinematic presentation. Figure treated older robots as repositories of proprietary hardware that could not safely enter a conventional resale or recycling channel.

That choice exposes a problem the humanoid robotics industry has barely started discussing. Companies are racing to build fleets, yet few publicly explain how owners will maintain, upgrade, resell, recycle, or securely destroy those machines.

Figure says its growing F.03 fleet made continued F.02 maintenance impractical. The company also says it is already developing F.04, increasing the pressure to move engineers away from obsolete hardware.

The furnace therefore represented more than an elaborate marketing stunt. It marked a rapid transition between hardware generations while raising unresolved questions about cost, security, sustainability, and customer support.

Figure AI Robot Decommissioning Became a Real Furnace Operation

Figure destroyed most of its F.02 fleet because the machines had become both obsolete equipment and containers for protected technology.

Figure published its decommissioning account on September 30, 2026. It said F.02 had supported its first BMW deployment, early household work, logistics experiments, and development of its Helix AI system.

Those achievements did not make the fleet worth maintaining. According to Figure, expanding F.03 production changed the calculation around engineering time, storage, spare parts, and ongoing support.

The company says each F.02 contained custom actuators and other proprietary components. Selling complete machines or handing them to an ordinary recycler risked exposing that hardware outside Figure.

Manual disassembly offered greater control, but it created a different cost. Figure said the work would occupy technical employees and delay the launch of F.04.

Adcock then asked social media users to propose a proper farewell. Schwarzenegger answered with three direct words: “You should melt them.”

Figure contacted the actor, who agreed to participate in the resulting film. His involvement transformed an internal disposal problem into a recognizable Terminator reference with immediate online appeal.

Executing the concept proved harder than proposing it. Figure says foundries in the United States and Mexico rejected robots containing lithium-ion batteries because those batteries could damage expensive equipment or create additional hazards.

The company eventually found a willing facility in Imatra. Its electric arc furnace uses three large graphite electrodes to melt steel scrap and can hold a 75-ton batch.

Figure received a 24-hour operating window covering six melts. Each cycle provided about 20 minutes before the steel cooled enough to form a hard surface crust.

The robots also needed a repeatable way to enter the furnace. Workers could have lowered every unit with industrial equipment, but Figure chose an autonomous jump for the main sequence.

Training started at its San Jose campus. The team placed large airbags below a second-floor platform and used movements from professional stunt performers as reference data.

Engineers then trained a new control policy in simulation. A control policy is the learned system that converts sensor inputs into coordinated movements across the robot’s joints.

Figure says the resulting policy guided the robots toward the furnace opening at a facility they had never previously visited. The units then executed the planned movement amid high temperatures and electromagnetic interference.

The original reporting describes camera equipment and other electronics failing around the furnace. Figure says the robot policies continued operating under those conditions.

One F.02 later recreated the Terminator 2 farewell by holding up a thumb while descending toward the metal. The scene emphasized the company’s cinematic goal rather than adding evidence about normal industrial performance.

Most of the F.02 fleet was destroyed during the operation. Figure says only a few examples remain in storage at its headquarters.

The material did not simply disappear. Figure shipped the resulting steel bars back to the United States and began machining them into a limited series of commemorative artifacts.

That recycling step gives the spectacle a circular ending, although Figure has not published a detailed material accounting. The public record does not quantify recovered steel, battery handling, emissions, slag, or unrecoverable components.

Those missing details matter because melting a complete robot is not equivalent to carefully recovering every valuable material. It solved Figure’s immediate security and staffing problem, but it did not establish a general recycling model.

Why Figure Retired F.02 While Its Successor Was Scaling

The retirement reflects Figure’s unusually fast hardware cycle, where a celebrated platform can become an engineering burden within one generation.

F.02 was not merely a laboratory shell. It gave Figure an operational platform for collecting data, testing physical tasks, and learning where humanoid hardware failed outside controlled demonstrations.

Its BMW work was particularly important. Factory deployments expose robots to repetitive schedules, tight workspaces, changing objects, and the reliability expectations of an established manufacturer.

Figure later moved the successor platform into the same industrial setting. The company announced that F.03 had arrived at BMW’s Spartanburg plant after F.02’s earlier assembly-line deployment.

That handoff helps explain the timing of the Figure AI robot decommissioning. The company was not retiring an isolated experiment while pausing development. It was reallocating resources toward a growing replacement fleet.

In April 2026, Figure said its BotQ manufacturing operation had delivered more than 350 F.03 robots. It also claimed a production cycle of one robot per hour, compared with one per day earlier.

The company characterized that result as a 24-fold throughput improvement achieved in less than 120 days. These are company-reported manufacturing figures, not independently audited production statistics.

Even so, they describe the pressure facing the older fleet. Every additional F.03 creates demands for testing, software deployment, maintenance procedures, inventory, and operational support.

Keeping F.02 active would require a parallel support structure. Engineers would need to maintain older actuators, diagnose generation-specific failures, and preserve tooling that newer hardware no longer uses.

Software also complicates the transition. F.03 was designed around later versions of Helix, Figure’s vision-language-action system for turning visual observations and instructions into robot behavior.

A vision-language-action model connects perception, language, and physical control. It must work through the sensors, hands, joints, and computing hardware available on a particular robot.

Hardware changes can therefore reshape the entire software stack. A policy developed for one hand design or sensor layout does not automatically transfer to another generation without testing.

Figure describes F.03 as having redesigned hands, a revised sensory system, softer exterior materials, improved audio, wireless charging, and additional battery safety measures. Those changes affect both capabilities and maintenance.

Supporting two generations would not only mean stocking two sets of parts. It could require separate validation, safety analysis, training data, troubleshooting procedures, and customer documentation.

The decision resembles an accelerated version of an old enterprise technology problem. A vendor must decide when maintaining a legacy platform costs more than migrating work to its replacement.

Humanoid robots make that calculation harder because the legacy system is physical. It contains batteries, motors, processors, structural components, sensors, and software credentials that require different retirement procedures.

A cloud service can remove old instances and revoke access keys. A humanoid cannot be securely deleted with one command, especially when its mechanical design remains valuable intellectual property.

That makes Figure’s choice understandable from an internal resource perspective. A small group of specialist engineers can be more valuable on F.04 than on weeks of controlled F.02 disassembly.

However, this reasoning works differently for future customers. A manufacturer operating purchased robots cannot treat every generational transition as an opportunity for a theatrical furnace trip.

Industrial buyers expect predictable service periods, replacement parts, repair options, data export, and safe disposal. Those expectations will grow as deployments move from pilots to long-term operating fleets.

Figure has explained why F.02 no longer served its development program. It has not yet published a broader lifecycle policy covering customer-owned or customer-operated robots.

That gap is the real pressure point. The Figure 02 molten steel sequence ended one internal fleet, while highlighting questions that commercial contracts will eventually need to answer.

The Terminator Sendoff Turned Disposal Into Marketing

Figure converted a difficult asset-retirement decision into a global demonstration, but visibility should not be confused with operational proof.

The project was built for attention. It combined human-shaped machines, real molten steel, autonomous movement, an iconic actor, and one of cinema’s best-known robot deaths.

Figure also anticipated skepticism created by synthetic media. The company emphasized that it trained real robots, shipped them to Finland, and filmed their physical jumps into the furnace.

That clarification served two purposes. It authenticated the spectacle while presenting the retirement process as another demonstration of Figure’s autonomy software.

The jump required coordinated balance, timing, and trajectory control. The robots reportedly performed it in an unfamiliar location despite conditions that interfered with surrounding electronics.

Those facts make the exercise technically interesting. They do not show that F.02 could navigate an active foundry, identify hazards, or perform a useful production task without supervision.

The system was trained for one narrow movement toward a known target. Figure’s team also controlled the scene, prepared the route, selected the timing, and worked within a scheduled furnace operation.

Independent reporting made the same distinction. A filmed autonomous maneuver shows a bounded capability, not general readiness for arbitrary work around molten metal.

That difference matters because humanoid robotics companies increasingly publish videos as evidence of progress. Carefully produced demonstrations can reveal genuine advances while omitting intervention rates, failed attempts, maintenance, and preparation time.

Figure disclosed several constraints, including the 24-hour window, six melts, simulated training, and practice jumps onto airbags. That context makes the video more informative than an unexplained social clip.

However, the company did not publish the number of practice attempts, policy failures, human interventions, or robots that completed each version of the maneuver. No independent evaluator monitored the operation.

The correct conclusion is therefore narrow. Figure says its F.02 robots executed trained autonomous jumps under difficult environmental conditions, and multiple reports document the physical furnace project.

The event does not prove broad foundry autonomy. It also does not validate F.03 reliability, F.04 development speed, or the economics of deploying humanoids at scale.

The marketing value may still be substantial. Robot companies compete for engineers, investors, manufacturing partners, and public attention long before deployment data becomes abundant.

A memorable retirement can keep Figure visible while the company transitions between products. It also turns obsolete hardware into a narrative about speed rather than waste.

That narrative says Figure is moving so quickly that yesterday’s breakthrough has already become expendable. It reinforces the image of a company willing to abandon working systems to accelerate the next generation.

The same message creates risk. Customers might interpret rapid retirement as evidence that expensive humanoid platforms will have short support lives or limited upgrade paths.

Developers may ask whether software investments survive a hardware transition. Factory operators may worry about spare parts, maintenance training, and tasks built around a discontinued design.

Sustainability critics can raise a different objection. Destroying functional machines for a promotional film looks wasteful unless the company clearly accounts for recovered materials and hazardous components.

Figure says the steel became commemorative artifacts. That is a form of reuse, but it does not resolve questions about batteries, electronics, rare materials, or total energy consumption.

Online reactions reflected both sides of this tension. Some viewers praised the practical filmmaking and Terminator reference, while others questioned why usable robots were not donated or dismantled.

Donation was probably incompatible with Figure’s intellectual-property concerns. It also could have created ongoing safety, software, and support obligations for machines never designed as open research platforms.

Yet this is precisely why public lifecycle planning matters. If a robot cannot be transferred, repaired externally, or economically dismantled, secure destruction becomes the default exit.

The viral film made that outcome entertaining. The underlying ownership model is more consequential than the spectacle surrounding it.

Humanoid Robot Makers Now Face a Lifecycle Test

The real competition is no longer limited to capability and production speed, because buyers also need credible plans for every machine’s final years.

Figure competes in a crowded field that includes Tesla’s Optimus, Agility Robotics’ Digit, Apptronik’s Apollo, Boston Dynamics’ Atlas, and several Chinese manufacturers.

These companies follow different commercialization strategies. Some prioritize human-shaped general-purpose systems, while others target bounded warehouse or manufacturing tasks.

Agility Robotics has focused Digit on moving containers through logistics operations. Its design uses birdlike legs and simple grippers instead of closely copying a human body.

An Associated Press account described Digit as commercially operational in warehouse and industrial facilities. It also identified customers and partners across logistics and manufacturing.

Apptronik has pursued industrial evaluations with companies including Mercedes-Benz. Tesla continues developing Optimus within its own manufacturing environment, while Boston Dynamics is preparing Atlas for factory work.

Most public comparisons examine walking, manipulation, autonomy, production targets, or task speed. Few compare support periods, repairability, component reuse, secure erasure, and end-of-life processing.

That omission is manageable while fleets remain small. It becomes expensive when hundreds or thousands of machines operate across customer sites.

Humanoids combine several difficult waste streams in one mobile system. A single unit can contain lithium-ion batteries, structural metals, circuit boards, cameras, computers, lubricants, polymers, and proprietary actuators.

Different components require different treatment. Batteries introduce fire and handling risks, while electronics contain materials that should not simply enter ordinary steelmaking processes.

Secure retirement adds another layer. A robot may store facility maps, task history, images, credentials, operational logs, and learned policies alongside protected mechanical designs.

Owners need a defensible way to erase that information before resale or recycling. Vendors also need to prevent retired components from exposing trade secrets or enabling unsafe modifications.

Traditional industrial equipment already has established channels for resale, remanufacturing, parts harvesting, and scrap recovery. Humanoid systems complicate those channels through software dependence and integrated design.

Academic work on end-of-life products consistently treats disassembly as important for recycling, reuse, and remanufacturing. A disassembly review also identifies technical and organizational barriers that make the work difficult.

Figure encountered those barriers directly. Its robots were apparently difficult enough to dismantle that using internal specialists threatened another product schedule.

That fact should influence the design of future generations. Engineers can reduce retirement costs by making batteries accessible, labeling materials, separating sensitive modules, and documenting safe disassembly.

Modularity can also protect intellectual property. A vendor might remove a small number of proprietary actuator or compute modules while sending ordinary structures through established recycling channels.

Software support needs similar planning. Customers should know whether a discontinued robot can run existing tasks, receive security fixes, and export operational data after its successor arrives.

A credible lifecycle package would define service duration, parts availability, battery replacement, ownership of collected data, decommissioning procedures, and responsibility for transportation.

It would also separate confidential destruction from general recycling. Securely removing sensitive modules does not necessarily require destroying every reusable mechanical part.

Figure’s internal prototypes may not justify such a complete program. Early development fleets often contain rapidly changing components, experimental wiring, and undocumented modifications.

Still, F.02 performed work at a major manufacturer and helped build Figure’s public credibility. Its retirement offers an early view of the challenges that customers will face later.

Competitors now have an opportunity to differentiate through longevity. A company that documents repairability and end-of-life recovery could reduce risk for buyers evaluating multiyear deployments.

This issue reaches beyond environmental messaging. Better serviceability can lower downtime, preserve residual value, simplify insurance, and make total ownership costs easier to forecast.

The furnace solved Figure’s immediate problem in one dramatic operation. A mature humanoid market will require a repeatable process that works without Schwarzenegger, a film crew, or a willing Finnish foundry.

What the F.02 Retirement Leaves Unanswered

Three signals will show whether this episode becomes an isolated prototype farewell or the start of a serious robot lifecycle strategy.

The first signal is Figure’s support policy for F.03. The company has shown that it can increase production, but production alone does not reveal how long those machines will remain serviceable.

Buyers need specific answers about warranties, repair turnaround, battery replacement, spare parts, software compatibility, and end-of-support notices. Published commitments would strengthen Figure’s commercial case.

Their absence would leave customers dependent on private contract terms and continuing vendor support. That dependence becomes more significant when robots perform daily production tasks.

The second signal is the design of F.04. Figure cited potential delays to that program as a reason for avoiding manual F.02 disassembly.

That experience gives the company a practical incentive to make its next platform easier to retire. Accessible batteries, removable compute modules, and separable proprietary hardware would reduce future labor.

Figure has not publicly detailed F.04’s architecture or release schedule. When information arrives, repairability and lifecycle design deserve attention alongside dexterity and autonomy.

A modular F.04 would suggest that the company learned from F.02’s disposal problem. Another tightly integrated platform could repeat the same conflict at a larger scale.

The third signal is operating evidence from F.03 deployments. Figure says F.03 has entered BMW’s Spartanburg plant and that production has expanded significantly.

The next meaningful numbers are not dramatic movement demonstrations. They are productive hours, intervention frequency, maintenance time, task completion, safety incidents, and component replacement rates.

Those measurements determine whether fast hardware turnover is productive iteration or an expensive cycle of replacement. They also reveal which components reach end of life first.

Competitors face the same test. Agility, Apptronik, Tesla, and Boston Dynamics will need to show that their machines remain economically useful after the launch presentation ends.

A robot that performs one impressive task but requires frequent specialist intervention can struggle to justify deployment. A repairable machine with narrower abilities might deliver greater practical value.

The Figure 02 molten steel event also raises questions for customers purchasing physical AI systems. Procurement teams should evaluate the exit process before accepting a pilot.

Who owns the operational data when a robot leaves the site? Which party removes credentials, transports batteries, certifies destruction, and records recovered materials?

Can the machine be resold, upgraded, or reassigned? Does the customer receive tools to export task data and preserve workflows when the vendor changes hardware?

Those questions sound administrative beside a robot jumping into molten steel. They ultimately decide whether humanoid fleets become durable infrastructure or disposable experiments.

The company deserves credit for explaining the resource conflict behind its decision. Figure did not pretend that maintaining every historical platform was worthwhile.

Its account also provides unusual detail about training, foundry access, timing, and environmental interference. That disclosure makes the project easier to evaluate critically.

However, Figure’s environmental and autonomy claims remain company-supplied. The public materials do not provide independent validation or a full lifecycle assessment.

The proper interpretation is therefore balanced. The robots apparently performed a real, narrowly trained maneuver, and their steel entered a commemorative reuse process.

At the same time, the operation avoided more selective disassembly because that work demanded scarce technical labor. That is an engineering limitation, not merely a creative choice.

Figure AI robot decommissioning succeeded as a film and protected the company’s proprietary hardware. Its broader value depends on whether future robots become easier to support, upgrade, and retire.

For developers, the lesson is to treat hardware generations as temporary while preserving software, data, and operational knowledge across them. For buyers, lifecycle terms belong in the initial evaluation.

The next Figure robot video will probably show a more capable machine. The more important document may be far less cinematic: a service and retirement policy explaining what happens after that machine becomes obsolete.