Figure's Helix gives humanoid robots a two-speed brain

Two weeks after cutting ties with OpenAI, humanoid robot maker Figure just unveiled its next big move: Helix, a generalist Vision-Language-Action (VLA) model designed to tackle the messiest place on earth — the human home.

Industrial robots thrive in factories because every bolt and conveyor belt sits in the exact same spot. But inside a house, robots face a nightmare of crumpled laundry, delicate glassware, and scattered toys they have never encountered before. Helix wants to fix this without needing thousands of hours of hand-coded demonstrations.

Fast reflexes meet slow thinking

Until now, robotics faced a nasty tradeoff. Vision-language models (VLMs) have rich common sense, but they think too slowly. Low-level motor controllers are lightning fast, but completely blind to context.

Figure solved this with a decoupled System 1 and System 2 architecture.

System 2 is a 7-billion-parameter internet-pretrained VLM running onboard at 7 to 9 Hz. It handles the high-level semantic heavy lifting, listening to voice prompts and surveying the scene.

System 1 is an 80-million-parameter transformer-based motor policy running at a blazing 200 Hz. It translates System 2's semantic goals into fluid, continuous upper-body motions — coordinating the head, torso, wrists, and all ten individual fingers in real time.

In one demonstration, Figure showed two robots collaborating to put away unfamiliar groceries. When one robot moved a package, the partner robot's System 1 instantly adapted its reach at 200 Hz while keeping System 2 focused on the overall chore.

What makes this more impressive is data efficiency. Figure trained Helix on just roughly 500 hours of teleoperated data, using an automated hindsight labeling process. A single set of neural network weights handles opening refrigerators, sliding drawers, and handling unseen objects without any task-specific fine-tuning.

Best of all, Helix runs entirely onboard Figure's embedded low-power GPUs without leaning on the cloud.

Why it matters

We think Helix is a huge validation moment for Figure. Breaking up with OpenAI could have stalled their momentum, but Helix proves Figure can build serious full-stack AI brains in-house.

For anyone hoping to see useful home helpers, this is the missing link. You cannot expect homeowners to hire roboticists to reprogram a bot every time they buy a new coffee mug. Helix's zero-shot generalization means robots can simply hear natural language and immediately manipulate objects they have never touched before.

Architecturally, splitting high-level semantic reasoning from high-frequency continuous control looks like the right formula. Bypassing tokenized action spaces in favor of direct continuous upper-body control gives humanoids the dexterity they desperately need.

Of course, handling thousands of novel objects on embedded hardware is still a long way from reliably cooking dinner or folding fitted sheets without incident. But by dramatically lowering training data requirements to just 500 hours, Figure may have found a genuine path to scale physical intelligence.