US company Figure AI has introduced a system that controls a humanoid robot using a single neural network. A four-minute kitchen demonstration is meant to showcase its capabilities.

The combination of locomotion and manipulation has been one of robotics’ most persistent challenges for decades. When a robot lifts an object, its balance changes; when it takes a step, its reach shifts. Arms and legs continuously influence each other.

Traditional systems bypass this complexity by splitting locomotion and manipulation into separate controllers: walk, stop, stabilize, grasp, walk again. Existing humanoid demonstrations — such as jumping or dancing — are typically planned offline. If an object shifts or contact occurs differently than expected, the behavior often collapses, Figure AI notes.

A single neural network for full-body control

Helix 02 aims to solve this by using one unified learning system that simultaneously controls the entire body. It expands last year’s Helix model, which only handled the upper body, to include legs, torso, head, arms, and individual fingers.

In a demonstration, Figure AI shows a robot unloading and reloading a dishwasher — performing 61 consecutive actions over four minutes without human intervention. The robot closes a drawer using its hip and lifts the dishwasher door with its foot when its hands are occupied.

The company describes this as the longest and most complex autonomous task ever performed by a humanoid robot. However, no data is provided on error rates or the number of attempts required to capture the footage. It is also unclear how the robot would perform in a modified kitchen environment. The dishes used are made of plastic. Still, if the actions were indeed performed autonomously as claimed, this represents a significant step forward compared to earlier demonstrations.

Three-layer architecture replaces hand-written code

The technical foundation is a three-layer architecture. System 0 — a neural network with 10 million parameters — was trained on more than 1,000 hours of human motion data and operates at 1 kHz for rapid corrections. According to Figure AI, it replaces 109,504 lines of hand-written C++ code previously used for balance and coordination.

Training was conducted in simulation using more than 200,000 parallel environments, a standard approach for sim-to-real transfer. Above this sits System 1, which connects all sensors with all joints and operates at 200 Hz. System 2 forms the top layer, responsible for language understanding and task planning.

New sensors enable finer manipulation

The hardware is based on the recently unveiled Figure 03 robot. Palm-mounted cameras provide visual input when objects are blocked from the head camera’s view. Tactile sensors in the fingertips detect forces as small as three grams.

Additional demonstrations include unscrewing a bottle cap, extracting a single pill from a medication blister pack, dispensing 5 ml from a syringe, and sorting metal parts — reportedly sourced from the company’s own manufacturing facility.

Figure AI describes its results as “early-stage” and aims to eventually deploy humanoid robots in both households and workplaces. The original Helix system was introduced last year.

Conclusion:

While Helix 02 is still an early-stage system, Figure AI’s approach marks a meaningful shift toward unified, end-to-end control of humanoid robots. If the technology proves robust outside controlled environments, it could significantly accelerate the deployment of versatile robots in real-world household and industrial settings.