Figure's Helix improves humanoid navigation
- Figure’s January 27 Helix 02 update moved its humanoid control from upper-body-only actions to full-body autonomy — walking, balancing, and manipulating in one loop. - The key demo was a 4-minute dishwasher task on Figure 03 with no resets or human intervention, backed by 1,000+ hours of motion data. - This matters because humanoids usually split walking and handling into brittle separate controllers; Figure is arguing one learned policy can do both.
Humanoid navigation sounds simple until you ask a robot to do anything useful while moving. Walking across a room is one problem. Picking something up is another. Doing both at once — without stopping, re-planning, or falling over — is the part that has kept most humanoids looking impressive in demos but awkward in real work. Figure’s Helix 02 is interesting because it goes straight at that gap. On January 27, 2026, the company said its latest system can control the whole robot as one continuous behavior, not as a chain of separate walking and manipulation modes. ### What actually changed? Earlier versions of Helix were pitched around upper-body control. Helix 02 extends that to the full robot — arms, hands, torso, balance, and locomotion — using one learned visuomotor system that takes in onboard sensing and outputs actions across the body in real time. Basically, Figure is saying the robot no longer has to “walk there, stop, stabilize, then act.” It can keep adjusting the whole time. (figure.ai) ### Why is that the hard part? Because movement and manipulation fight each other. Lift a plate and the center of mass shifts. Step forward and the reach geometry changes. Turn your torso and the hand path changes too. Traditional humanoid stacks handle this with stitched-together controllers and state machines — one for locomotion, one for grasping, one for balance, then a scheduler deciding when to hand off. That works, but it is slow and brittle. If the object moves or contact happens differently, the routine can unravel. (figure.ai) ### What did Figure show? The headline demo was a Figure 03 robot unloading and reloading a dishwasher across a full kitchen for about 4 minutes, end to end, with no human intervention or resets. Figure called it the longest-horizon, most complex fully autonomous humanoid task it had shown so far. The point was not dishwashers. The point was continuous loco-manipulation — walking, reaching, carrying, stacking, and balancing as one behavior. (figure.ai) ### Is this only about kitchens? No — and that is where the navigation angle gets more concrete. In a March 9 update, Figure showed Helix 02 tidying a living room, side-stepping through narrow gaps between furniture while still handling objects, stowing tools under an arm mid-task, and reorienting objects in hand. That matters because homes and warehouses are messy in the same annoying way for robots — paths narrow, objects shift, and the robot has to keep working while moving. (figure.ai) ### What is doing the heavy lifting? Part of it is the model, but part of it is the robot body. Figure 03 was built around Helix, with a redesigned vision stack, lower latency, a wider field of view, palm cameras, and tactile sensing in the hands. Figure says those upgrades are meant to keep perception stable in cluttered spaces and during occlusion — like when an arm blocks the main camera while reaching into a cabinet. Better navigation here is really better perception-plus-control. (figure.ai) ### Does this connect to real work? Yes. Figure had already been pushing Helix in logistics, where it said package handling speed improved to 4.05 seconds per item from about 5.0 seconds, while barcode orientation success rose to roughly 95% from about 70%. That is a different task from room navigation, but it shows the same pitch: scale the data, improve the policy, and the robot gets less scripted and more adaptive. (figure.ai) ### So is navigation solved now? Not really. These are still company-run demos, and real deployments are harsher than curated rooms. But the shift is real. The field has spent years with humanoids that could either move well or manipulate well. Figure is trying to collapse those into one control problem. If that works outside demos, humanoids stop being statues with hands and start being mobile workers. ### Bottom line? (figure.ai) The news is not that Figure taught a robot to dodge obstacles. The news is that it is framing navigation as part of one full-body intelligence stack. That is the version that could actually matter in warehouses, factories, and eventually homes. (figure.ai)
