Figure says Helix 02 demonstrated full-body control in a four-minute dishwasher task, using a unified neural architecture to coordinate walking, manipulation and balance. In the company-produced demonstration, the robot uses onboard sensors and completes the sequence without human intervention.
The announcement, published on January 27, 2026, presents Helix 02 as an expansion of the earlier Helix system, which Figure says controlled a humanoid’s upper body from camera input. The new system is intended to control the robot’s entire body, including its legs, torso, arms, hands and balance.
Figure describes the demonstration as a capability showcase rather than a conventional product launch. The supplied announcement does not provide independent testing, a formal benchmark, commercial availability, production plans or safety certification.
Contents
What changed
The central change is the move from upper-body control to what Figure calls full-body autonomy. A humanoid performing a household or industrial task must do more than move its hands: it needs to walk to the work area, maintain balance while reaching, coordinate contact with objects and sometimes use parts of its body other than its hands.
Figure uses the term loco-manipulation for this combination of locomotion and object manipulation. In such tasks, movement and manipulation are interdependent. A step can change the robot’s balance, while pushing, pulling or lifting an object can affect its posture and foot placement.
In the dishwasher demonstration, Figure says Helix 02:
- walks to a dishwasher;
- unloads and stacks items;
- reloads the dishwasher;
- starts the dishwasher;
- completes the sequence as one continuous four-minute task.
The company says the sequence contains 61 loco-manipulation actions performed in the correct order, with implicit error recovery and task state maintained across several minutes. Figure also highlights whole-body actions such as closing a drawer with the robot’s hip and lifting the dishwasher door with its foot.
Figure presents the sequence as a demonstration that perception, task sequencing, balance and manipulation can operate together over a room-scale sequence. The supplied source does not state how many attempts were made, how many succeeded, whether the video was a single successful take or what environmental preparation was required. Its reliability and repeatability were not evaluated in the supplied evidence.
How the control system works
Helix 02 is organised into three stated layers, with different responsibilities and operating speeds.
System 2: goals and sequencing
System 2 handles scene understanding, language, semantic goals and task sequencing. It operates at a slower timescale than the lower-level control systems.
The announcement does not disclose how much of the dishwasher sequence was generated online, how much was preconfigured or what recovery policies were built into the demonstration. That distinction matters: a system that independently interprets changing conditions is different from one executing a tightly specified task under controlled conditions.
System 1: whole-body visuomotor control
System 1 converts sensory information and higher-level goals into full-body joint targets. Figure says it operates at 200 Hz, meaning it can update those targets 200 times per second.
Its inputs include:
- head cameras;
- palm cameras;
- fingertip tactile sensors;
- full-body proprioception.
Proprioception is sensing of the robot’s own body state, such as joint positions and movement variables. It helps a controller know where the robot’s limbs are and how they are moving, rather than relying only on external camera views.
System 1 controls the legs, torso, head, arms, wrists and individual fingers. Figure says the palm cameras and tactile sensors are capabilities from its Figure 03 hardware platform. The company also says the tactile sensors can detect forces as small as three grams.
Tactile sensing can provide contact information that vision alone may not provide when an object is occluded or when grip force must be regulated. The announcement does not provide independent measurements of tactile accuracy or show how performance changes when vision and touch disagree.
System 0: fast balance and contact control
System 0 is the lowest-level layer. Figure describes it as a 10-million-parameter neural network that receives the robot’s full-body joint state and base motion, then outputs joint-level actuator commands at 1 kHz.
This layer handles balance, contact and whole-body coordination at a faster timescale than System 1. A hierarchical control system of this kind separates slower semantic decisions from rapid physical corrections. System 2 can handle the goal and sequencing, System 1 can generate coordinated body motion and System 0 can make high-frequency adjustments as the robot’s body interacts with the environment.
Figure says System 0 was trained with more than 1,000 hours of joint-level retargeted human-motion data. It also says the network was trained entirely in simulation across more than 200,000 parallel environments using domain randomisation—a method that varies simulated conditions to reduce reliance on one precisely configured virtual environment.
Simulation training and domain randomisation are part of Figure’s reported method, but they do not by themselves establish how the controller performs on physical hardware. The announcement does not disclose the full network architecture, training objective, dataset composition or deployment hardware specifications.
What the demonstrations show
Alongside the dishwasher task, Figure reports four autonomous dexterity demonstrations:
- unscrewing a bottle cap;
- extracting a pill from a medicine organiser;
- dispensing exactly 5 ml from a syringe;
- picking small metal pieces from a cluttered box.
Figure says these videos were fully autonomous and not teleoperated. The company presents the tasks as demonstrations of the capabilities it attributes to Helix 02, including perception, tactile sensing, hand control and whole-body movement. The supplied announcement does not specify which sensors or control layers were used in each individual video.
The four tasks involve bottle-cap removal, pill extraction, measured syringe dispensing and picking small metal pieces from a cluttered box. These descriptions show the types of manipulation Figure is presenting, but they do not establish quantitative accuracy or reliability: the source does not report the number of trials, success rates, failure modes or repeatability.
Figure positions Helix 02 as a step towards general-purpose humanoid robots for homes and the workforce. The demonstrations may be relevant to industrial handling, sorting and other multi-step tasks, but they do not establish that the system is ready for unsupervised operation in factories, homes or environments with people.