Firefly liquid silver mercury

Why Figure AI trained its humanoid robots to self-destruct like the Terminator

Figure AI had a problem: what do you do with an obsolete fleet of humanoid robots that you really, really don’t want falling into the wrong hands?

The robots in question were its Figure 2s, recently superseded by the shiny new Figure 3. Figure had kept the new 3 (as its friends call it) model shrouded in secrecy: when it gave Time magazine a sneak peek ahead of its launch last fall, it was hidden behind locked security doors while its predecessor roamed freely around the building. Described by Figure as its first mass-producible humanoid, 3 represented a serious overhaul: smaller and stronger joints, components that are 90% cheaper to manufacture and less of a tendency to catch fire (always a plus).

Somewhat pitifully, the 2s were orchestrators of their own obsolescence, working alongside humans to harvest training data for Helix, the neural network that would run 3s. And what did they get as thanks? An instruction to eliminate their entire species. It’s hard out there for a ’bot.

Why so cruel, Figure? Well, the 2s contained “custom-built actuators and other intellectual property” and needed a disposal method that protected the proprietary hardware from another company reverse engineering it and achieving market domination.

And if that sounds familiar, you are correctly recalling the plot of Terminator 2: Judgment Day, in which a robot containing an advanced CPU and other futuristic technology needed to be disposed of before anyone could reverse engineer it and enable an AI bent on global domination.

So when CEO Brett Adcock asked the internet what to do with the doomed 2s, it was naturally Arnold Schwarzenegger who suggested melting them. In Terminator 2, the evil new T-1000 pursues the surviving trio of Arnie’s rehabilitated T-800 and John and Sarah Connor, cornering them in a steel mill—which frankly seems a little like the Wicked Witch of the West ambushing Dorothy in the shower, but, whatever; it was destroyed and the world survived.

It’s important to note that crushing them with a hydraulic press was not advised: the original Terminator had already tried that, and its surviving arm and CPU were subsequently reverse engineered by the fictional Cyberdyne Systems, demonstrating what happens when you don’t dispose of your obsolete tech securely enough.

Figure could have just driven the 2s to a forge on errand day and dropped them off, but where’s the fun in that? Instead, the company asked itself, how do you teach a humanoid robot to convincingly destroy itself in an arrestingly cinematic way?

Ironically, that was one thing the Terminator couldn’t do. “I cannot self-terminate,” the T-800 tells Sarah Connor, handing her the controls so she can lower him into the bubbling cauldron of molten steel.

Figure’s robots had no such qualms. They had the best training and were waiting for their moment to shine. Figure had brought in stunt performers, used their movements as reference data, built a new AI model in simulation and trained the 2s to jump accurately into a furnace at a foundry they’d never visited. Not just any furnace: a 75-ton electric arc furnace in Finland. Figure had only a 20-minute window during each melt before a crust formed on the steel, and the heat and electromagnetic interference were intense enough to knock out some of the camera equipment.

Despite the conditions, the robots’ AI policies kept running. They even emulated the T-800’s thumbs-up signal, the last thing we see before it disappears.

There’s something beautifully poetic about the whole thing. CGI was still in its infancy in 1991 and it took a team of up to 35 people to work on the T-1000’s five minutes of screen time, and rendering just 15 seconds could take up to 10 days. If you’re following along then, 35 years after those computers learned to make a fake robot move like a real person, AI learned to make a real robot move like the fake one. In other words, robots trained by people imitated robots played by people pretending to be robots. Glad that’s clear.

As Sarah Connor drives down a highway with John at the end of T2, she reflects on her renewed hope for an unknown future, musing, “If a machine, a Terminator, can learn the value of human life, maybe we can too.” The Terminators came from 2029, incidentally, which is now barely more than two years away. Everyone feeling good about the timeline? Great.

Elly Trickett

Editorial Director, Inbound & News

IBM

Related solutions
IBM Terraform

IBM Terraform® provides organizations with a single workflow to provision their cloud, private data center and SaaS infrastructure and continuously manage it throughout its lifecycle.

Explore IBM Terraform
Automation solutions

Automate provisioning, networking, identity governance and app workflows across hybrid cloud with AI and IaC.

Explore automation solutions
Cloud consulting services

Unlock the full potential of hybrid cloud in the era of agentic AI.

Explore cloud consulting services
Take the next step

Discover how IBM Terraform® and IT automation solutions combine automation and real-time observability to boost resilience and accelerate growth.

  1. Discover IBM Terraform
  2. Discover automation solutions
Related solutions
IBM Terraform

IBM Terraform® provides organizations with a single workflow to provision their cloud, private data center and SaaS infrastructure and continuously manage it throughout its lifecycle.

Explore IBM Terraform
Automation solutions

Automate provisioning, networking, identity governance and app workflows across hybrid cloud with AI and IaC.

Explore automation solutions
Cloud consulting services

Unlock the full potential of hybrid cloud in the era of agentic AI.

Explore cloud consulting services
Take the next step

Discover how IBM Terraform® and IT automation solutions combine automation and real-time observability to boost resilience and accelerate growth.

  1. Discover IBM Terraform
  2. Discover automation solutions