Figure’s Helix: How One AI Model Is Teaching Humanoids to Work

Posted on by IE Contributor

For most of robotics history, teaching a machine a new skill meant one of two expensive things: an expert writing code line by line, or thousands of human demonstrations for a single task. Figure AI is betting on a third path. Its AI system, Helix, aims to let a humanoid robot hear a plain-language request, look at the scene in front of it, and work out the motions on its own. In under two years, Helix has gone from controlling a robot’s arms and fingers to steering its entire body through homes it has never seen. Here is what Helix is, how it has evolved, and what it means for businesses watching the humanoid race.

What Helix Is: A “Vision-Language-Action” Brain

Figure introduced Helix in February 2025 as a generalist vision-language-action (VLA) model, a type of AI that combines what a robot sees, what it is told, and what it physically does into a single learned system. In its launch write-up, the company described a two-part design borrowed loosely from psychology’s “fast and slow thinking”:

  • System 2, a 7-billion-parameter vision-language model that runs at roughly 7 to 9 times per second, interprets the scene and the spoken instruction.
  • System 1, a much smaller 80-million-parameter network, turns that understanding into continuous motor commands 200 times per second.

The split solves a long-standing tradeoff. Large language-and-vision models are broad but slow; classic robot control policies are fast but narrow. By having the “slow” model hand a compact goal signal to the “fast” one, Helix could, according to Figure, coordinate 35 degrees of freedom across the wrists, fingers, torso and head of its Figure 02 robot. Figure said the whole model ran on low-power GPUs inside the robot, trained on about 500 hours of teleoperated data, and could pick up thousands of household objects it had never encountered simply by being asked, including a toy cactus when told to grab “the desert item.”

Independent coverage was enthusiastic but measured. TechCrunch noted that VLAs were a young category, with Google DeepMind’s RT-2 the best-known earlier example, and cautioned that polished demo videos hide a lot of behind-the-scenes work. That caution is worth keeping in mind for every milestone that followed.

Helix 02: From Arms to the Whole Body

The first Helix controlled only the upper body. In January 2026, Figure released Helix 02, which extends control to legs, balance and locomotion, so the robot can walk and manipulate objects as one continuous behavior. Roboticists call this “loco-manipulation,” and it is notoriously hard: picking something up shifts your balance, and taking a step changes your reach.

Helix 02 adds a third layer, System 0, a 10-million-parameter controller running at 1,000 times per second that handles balance and contact. Figure says it was trained on more than 1,000 hours of human motion data in simulation, and that it replaced 109,504 lines of hand-written C++ code. The company’s headline demo was a four-minute, fully autonomous dishwasher task across a full-sized kitchen, involving 61 sequenced actions, with the robot nudging a drawer shut with its hip and lifting the dishwasher door with its foot when its hands were full.

Helix 02 also tapped new hardware in Figure 03, the company’s third-generation robot: palm cameras and fingertip sensors that Figure says can detect forces as small as three grams. With them, the company showed the robot unscrewing a bottle cap, picking a single pill from an organizer and pushing exactly 5 ml from a syringe.

Helix 2.5: Learning From People, Working in Strangers’ Homes

The most recent step, Helix 2.5, announced in September 2026, tackles the biggest obstacle to household robots: every home is different. Earlier systems mostly learned in the same places they would later work. For Helix 2.5, Figure pretrained the model on Index, its large dataset of people filmed doing everyday tasks, then adapted it to three chores: tidying toys, folding towels and making beds. It then tested the robot in 30 rented Bay Area homes where no training data had been collected.

The company’s key claims:

  • Pretraining on human video lifted full-task success from 9% to 56%, with everything else held constant.
  • Helix 2.5 needed half as much task-specific robot data as a comparable Helix 02 behavior.
  • Doubling the amount of human pretraining data improved results in a predictable way, which Figure describes as a human-to-robot “scaling law” similar to those that guided large language model development.

The fine print matters. Humanoids Daily’s breakdown of Figure’s own charts shows 237 successful trials out of 420: 67% for bed making, 62% for towel folding and just 40% for tidying toys. “Zero-shot” applies to the homes and objects, not to the chores themselves, which were trained elsewhere. The outlet also noted that this is a company-run evaluation, and that a rival founder, Sunday Robotics’ Tony Zhao, publicly argued that useful work requires reliability as well as generalization. Figure itself writes that general humanoid robotics “is not solved.”

Where Helix Is Already Working: The Factory Floor

While the home is Helix’s long-term target, industry is where it is earning its keep today. According to a BMW Group press release from June 2026, a Figure 02 robot supported production of more than 30,000 BMW X3 vehicles over ten months at Plant Spartanburg in South Carolina, inserting sheet-metal parts for welding. BMW is now moving Figure 03 into a logistics “sequencing” job: picking unsorted parts from bins, sorting them into trolleys, and moving those trolleys to a collection point for delivery to the assembly line.

BMW frames humanoids as a complement to existing automation, aimed at monotonous, ergonomically demanding or safety-critical tasks rather than wholesale replacement of workers. That framing is a useful reality check on the hype: the near-term value is in narrow, repetitive jobs that are awkward for fixed industrial robots.

Practical Takeaways

  • Language is becoming the programming interface. Helix’s core promise is that new robot behaviors can be requested in words rather than coded. Teams that document processes clearly in plain language will be better placed to use such systems.
  • Data, not hardware, is the moat. Figure’s biggest gains came from more and broader training data. Expect competition among robotics firms to center on who can gather the most useful real-world experience.
  • Read the success rate, not the video. A 56% full-task completion rate in unfamiliar homes is a real research advance and nowhere near appliance-level reliability. Ask vendors for trial counts, failure definitions and intervention rates.
  • Industrial pilots come first. Structured, repetitive jobs such as parts sequencing are where humanoids are proving value now; household robots remain a longer bet.
  • Watch the scaling claims. If Figure’s human-to-robot scaling law holds up in independent testing, progress could accelerate quickly with more data and compute.

Conclusion

Helix captures where robotics is heading: away from hand-coded routines and toward a single learned “brain” that connects seeing, understanding and acting. In roughly 20 months, Figure has moved it from grabbing objects on a table to walking through strangers’ homes making beds. The results are still company-reported, and reliability is the gap that separates impressive demos from dependable products. But between real factory work at BMW and steadily improving generalization, Helix is one of the clearest signals yet that general-purpose humanoid robots are moving from science fiction toward the workplace.

Sources