Figure AI demo Helix 02 cleans living room

- Figure AI published a March 9 demo of Helix 02 directing its Figure 03 humanoid to tidy a living room end to end. - The company says one learned policy now runs full-body control from pixels, with 1,000+ hours of human motion data behind it. - It matters because humanoid control is shifting from hand-coded stacks toward learned, room-scale autonomy in messy homes.

Humanoid robots usually look impressive right up until they have to do something ordinary. A living room is the hard version of ordinary — soft objects, clutter, awkward reaches, changing layouts, and no fixed workcell. That is why Figure’s new Helix 02 demo matters. On March 9, the company showed its Figure 03 robot tidying a living room on its own, using the same full-body AI stack it introduced a month earlier for kitchen cleanup. ### Why is a living room harder than a factory? A factory gives you repeatability. Parts arrive in known places, lighting is controlled, and the robot can repeat the same motion all day. A living room is the opposite — toys on the floor, cushions out of place, objects partly hidden, and surfaces that need different kinds of contact. The robot has to walk, balance, identify objects, decide where they belong, and manipulate them without a reset between steps. (figure.ai) That is exactly the class of problem humanoid companies say they want to solve, because homes are messy in the same way the real world is messy. ### What did Figure actually show? The demo shows the robot moving through a room, picking up scattered items, straightening couch cushions, and wiping surfaces at roughly human pace. Figure framed it as a continuation of the dishwasher demo from the Helix 02 launch, but in a different room with different object types and more soft, deformable interactions. The point was not one flashy grasp. (figure.ai) The point was a chained sequence of actions across an entire room. ### What is Helix 02, really? Basically, it is Figure’s second-generation vision-language-action system for humanoids. The company says Helix 02 controls the full robot directly from onboard sensing, tying perception, reasoning, walking, manipulation, and balance into one system rather than splitting them into separate hand-built modules. Figure describes it as “all sensors in, all actuators out,” which is the important architectural claim here — one learned policy spanning the whole body. (figure.ai) ### Why do people keep mentioning 1,000 hours? Because Figure says the whole-body controller, called System 0 inside the Helix 02 stack, was trained on more than 1,000 hours of human motion data plus sim-to-real reinforcement learning. That matters because walking and manipulation are usually trained or engineered separately. Figure is arguing that human demonstrations can teach the robot coordinated, human-like movement patterns across the entire body, then reinforcement learning helps make those patterns stable and useful on the real machine. (figure.ai) ### What changed versus the first Helix? The original Helix pitch was about upper-body generalist control and multi-robot collaboration. Helix 02 extends that to loco-manipulation — walking while reaching, balancing while carrying, and handling long-horizon tasks across a room. In plain English, Figure moved from “the robot can use its arms intelligently” to “the robot can do a room-scale chore without breaking the task into separate robot brains.” That is a bigger jump than it sounds. (figure.ai) ### Is this just a polished demo? Partly, yes — every robotics video is selected to show progress. But the more interesting signal is the pace of iteration. Figure went from Helix 02’s kitchen launch in early 2026 to a living room demo in March, and then to a two-robot bedroom reset demo published on May 8. That sequence suggests the company is trying to prove generalization across rooms and task types, not just one benchmark scene. (figure.ai) ### What is the real engineering bet? The bet is that learned full-body control can replace a lot of brittle robotics software. Traditional stacks are easier to inspect, but they break when the environment stops cooperating. Learned policies are messier to validate, but they can adapt to novel objects and layouts. The catch is safety — once one policy controls the whole body, failures can couple together. (figure.ai) A bad perception judgment is no longer just a bad grasp. It can become a balance problem too. ### Bottom line? Figure’s living room demo does not prove home robots are ready for your apartment. But it does show where the field is heading — away from scripted tricks and toward learned, room-scale autonomy. If that trend holds, the real milestone will not be one robot folding one shirt. It will be a humanoid that can enter a messy room and just keep going. (figure.ai)

Get your own daily briefing

Scout delivers personalized news, insights, and conversations tailored to your role and industry.

Download on the App Store

Shared from Scout - Be the smartest in the room.