Robot smoothly does household chores! Figure showcases neural network achievements, claimed to be "completely autonomous"

On March 10th, local time, humanoid robot company Figure released a new demonstration video. Figure states that its robot, Figure 03, equipped with the new AI system Helix 02, can fully autonomously and end-to-end organize a living room.

The two-and-a-half-minute video shows improvements in the robot’s object recognition, movement, and hand control capabilities.

The scene is an open-plan living room with an adjacent open kitchen. The robot enters holding a spray bottle and cleaning cloth, bends over, squats, moves a cup aside, sprays disinfectant and wipes the table, continues tidying the surface and sofa, and stores toys on the table and sofa. Finally, it presses a button to turn off the TV and “meticulously” straightens the remote control (despite not gently placing it on the table). It is observed that the robot handles objects differently—carefully picking up cups and tossing sofa cushions from its hand—indicating it can correctly recognize different objects.

Since its inception, Figure has focused on home use scenarios, aiming to develop versatile humanoid robots that assist humans at home and at work. Previously, Figure released videos of robots performing simple chores like washing dishes and folding clothes.

In fact, household chores are extremely challenging for robots. Unlike standardized factory tasks, home environments are unpredictable, with cluttered arrangements, furniture obstructing movement, and soft items like towels and pillows placed variably. Many actions require both hands, and some tasks demand freeing one hand during operation. Almost all movements involve moving within a room while manipulating objects.

How does Figure achieve this?

Figure states that it does not rely on writing separate “instruction manuals” for each action. Instead, it learns through a universal “brain.” This brain directly interprets the visual input from the camera (pixels) and controls every part of the body (full-body control) to perform complex sequences of actions. These new capabilities do not require reprogramming; simply providing more examples of new household tasks allows the robot to learn on its own.

According to the company, Helix 02 is currently its most powerful AI system, launched on January 27, 2026, and can be regarded as the “intelligent brain” of the Figure robot. Its main feature is that it is a single neural network that does not split control modules and can directly drive full-body movements through pixel input. This visual-language-action (VLA) model enables the robot to see with its eyes, understand language through listening, and coordinate full-body actions under the guidance of the same “brain” to complete complex tasks.

Real-world data is crucial for training the Figure robot.

Previously, Figure announced a “big project” (Project Go-Big) to build the world’s largest pre-training dataset for humanoid robots. The company is also collaborating with Brookfield, which owns over 100,000 residential units, to accelerate this process.

(Source: Cailian Press)