Figure AI's Helix VLA replaces hand-coded visuomotor logic with end-to-end neural control

Humanoid robotics has had a coordination problem for years. Walking works. Grasping works. But doing both at once — while the world shifts underneath you — is where a lot of robots turn awkward, brittle, or fully stuck. Figure AI’s latest move is to stop stitching those behaviors together with hand-coded logic and let a neural stack run the whole body instead. On January 27, 2026, the company introduced Helix 02 and said it had replaced 109,504 lines of control C++ with a learned whole-body controller on its Figure 03 robot. (figure.ai) ### What actually changed? The big change is architectural. Figure’s original Helix, released on February 20, 2025, was pitched as a vision-language-action model for upper-body control — basically seeing, understanding language, and moving arms, hands, torso, head, and fingers in one learned system. Helix 02 extends that idea to the full robot, so walking, balancing, reaching, carr(figure.ai)us control problem instead of separate modules. (figure.ai) ### Why were the old controllers a problem? Traditional humanoid stacks usually decompose behavior into steps: walk, stop, stabilize, reach, grasp, walk again. That sounds sensible, but the catch is that real motion does not come in neat turns. If a robot lifts a bowl, its balance changes. If it steps sideways, the arm trajectory changes. If a cabinet door swings differently than expe(figure.ai)rgument is that these controller handoffs are the brittle part — not just the individual skills. (figure.ai) ### So what is Helix 02 made of? Figure now describes three layers. System 2 handles slower scene understanding and language-level intent. System 1 turns perception and intent into high-rate motion targets. System 0 is the new piece — a learned whole-body controller that executes balance, contact, and coordination across the robot. The company says System 0 was trained on more than(figure.ai) reinforcement learning. (figure.ai) ### Why does deleting code matter? Because this is really a claim about where robot capability should come from. Instead of engineers writing ever more edge-case logic for balance and coordination, Figure wants those priors learned from data. The company says System 0 alone replaced 109,504 lines of hand-engineered C++, which is less a software cleanup story than a “Software 2.0” b(figure.ai)hen branches, more learned reflexes. (figure.ai) ### What did the robot actually do? Figure’s showcase was a four-minute kitchen task on Figure 03: walk to a dishwasher, unload items, move through the room, place objects in cabinets, reload dishes, and start the machine again — all from onboard sensing, with no resets or human intervention. Figure called it the longest-horizon, most complex autonomous task yet completed by a huma(figure.ai)emo matters because it bundles locomotion, dexterity, and balance into one uninterrupted behavior. (figure.ai) ### Why does Figure 03 matter here? Helix 02 is not just a software story. Figure 03, introduced on October 9, 2025, was built around faster vision, lower latency, wider field of view, embedded palm cameras, and fingertip tactile sensing down to 3 grams of pressure. Those sensors give the neural controller denser feedback, especially when the main cameras are occluded during close (figure.ai)rol needs better robot senses. (figure.ai) ### Is this bigger than one demo? Yes. The broader point is that home and workplace robots will run into too many object types, layouts, and weird edge cases for manual scripting to scale. Figure made that case with the first Helix release in 2025, and Helix 02 pushes it further by treating the whole body as one learned policy. If that approach holds up outside curated demos, it changes the engineering playbook for humanoids. (figure.ai) ### Bottom line Figure is trying to replace brittle robot choreography with learned whole-body control. That does not mean the hard part is solved. But it does mark a real shift — from robots executing hand-authored sequences to robots learning the coordination itself. (figure.ai)