Figure's Helix 02 tackles living room clutter

Tidying up a living room sounds simple, until you actually try to teach a robot how to do it.

Unlike an industrial assembly line, living rooms are pure chaos. Pillows deform, towels bunch up, coffee tables leave impossibly tight gaps, and objects constantly need to be held, passed, and reoriented while walking. It is a nightmare for traditional robotics.

Figure just demonstrated its Helix 02 robot tackling living room chores, and the results look remarkably natural. Building on a kitchen cleanup test released about a month ago, the humanoid navigated an entire living room using a single end-to-end neural system running directly from pixels to whole-body motor controls.

Casual tool use and dynamic towel whips

What makes the new footage stand out is not just what the robot moves, but how it moves.

Helix 02 picks up a spray bottle to wet a dirty surface, then uses a towel in the other hand to wipe down the mess with noticeable force. When it needs both hands free, it whips the towel over its shoulder or stows it under its arm while walking.

For cleaning up toys, it picks up a bin with both hands, holding it steady while scooping wooden blocks off a table. To move faster, it tucks the bin under one arm to free both hands for picking up clutter from the floor.

It even picks up a TV remote, flips and reorients it entirely within its hand, and presses the power button to turn off the screen. When it sees an out-of-place throw pillow, it tosses it back onto the couch with a smooth, dynamic flick.

All of this happens while the robot side-steps through narrow clearances between a couch and coffee table, keeping its hands working without pausing its stride.

Zero new algorithms, just more data

The most important detail here is under the hood: Figure claims it did not write a single new algorithm or build any custom controller for these tasks. Helix learned the entire living room routine simply by being fed more demonstration data into its existing model.

Traditionally, a demo like this would require chaining together separate modules for navigation, inverse kinematics, object classification, and grip planning. Helix collapses all of that into one unified model that looks at raw video pixels and outputs coordinated full-body motion.

Because the system is fully learned, adding new skills looks much more like training a language model than writing robotic software.

Why it matters

We think this demo is a major validation of the end-to-end learning approach for humanoids.

Watching a robot naturally tuck a container under its arm or whip a rag onto its shoulder shows an emerging sense of whole-body intuition. In classical robotics, chaining those micro-decisions together would break down the second an object slipped by two inches. Here, the system reacts to continuous visual feedback without robotic hesitation.

For the robotics race, Figure is putting immense pressure on competitors relying on modular stacks. If a single neural architecture can seamlessly move from kitchen tasks to living room tidying just by soaking up more data, the development timeline for general-purpose chores shrinks dramatically.

Still, it pays to stay grounded. Figure has not shared success rates across dozens of consecutive runs, and there is a massive gap between a controlled set-piece living room and a home with running pets and unpredictable kids. But as a proof of concept for scalable humanoid intelligence, it looks uncomfortably close to real home help.