One policy stack for movement and manipulation
Figure introduced Helix 02 on January 27, 2026 as a control system for its Figure 03 humanoid. The company demonstrated a continuous four-minute task in which the robot walked around a kitchen, unloaded dishes, placed them in cabinets, reloaded the dishwasher and started it. Figure states that the run was autonomous, used onboard sensors and had no resets or human intervention.
The technical goal is whole-body loco-manipulation. Walking changes a robot's balance, reach and camera view; carrying an object changes its mass distribution; opening a door creates contact forces that travel through the arms and feet. Conventional systems often join separately designed walking, standing and manipulation controllers with state machines. Helix 02 instead connects a learned hierarchy from visual input to all of the robot's joints.
Three systems at different speeds
The slowest layer, System 2, interprets the scene and language goal and sequences semantic behaviors. System 1 is a visuomotor transformer running at 200 hertz. It receives head-camera images, palm-camera images, fingertip forces and proprioception, then produces targets for the legs, torso, head, arms, wrists and fingers.
System 0 runs at 1 kilohertz and converts those targets into coordinated, stable movement. Figure describes it as a 10-million-parameter policy trained in simulation across more than 200,000 parallel environments and on more than 1,000 hours of retargeted human-motion data. Simulation randomization is intended to make the controller tolerate differences between virtual dynamics and the physical robot.
This hierarchy does not mean one network performs every calculation at the same rate. It connects slow task reasoning with a faster perceptual policy and an even faster balance-and-contact controller. That division is a practical response to the latency mismatch between language models and motor control.
Hardware closes perception gaps
Figure 03 adds cameras in the palms for cases where the head camera cannot see an object inside a cabinet or behind the hand. Fingertip sensors are claimed to detect forces as small as three grams, allowing the policy to respond to contact and slip. In separate demonstrations, the system unscrewed a cap, manipulated individual pills and dispensed a requested syringe volume.
During the dishwasher sequence, the company counted 61 loco-manipulation actions. The robot carried objects while walking, used both hands and sometimes used its hip or foot when its hands were occupied. These behaviors illustrate coupled control, but a count of actions is not a standardized measure of general intelligence.
Evidence from a curated demonstration
Figure provides detailed architecture, frequencies, sensor inputs and training-scale figures. It does not provide model weights, full evaluation data, repeated-trial success rates or an independent reproduction. The central evidence is company-produced video and description. A continuous four-minute run is stronger than a montage of disconnected movements, but it cannot reveal how often attempts fail or which scene conditions were selected.
Tasks involving pills and syringes are manipulation demonstrations, not authorization for medical use. Likewise, success in a prepared kitchen does not establish safe operation in arbitrary homes with children, pets, clutter or unexpected contact. The company presents the result as a robotics milestone, and its broader claims should be evaluated as vendor claims until outside testing is available.
Helix 02 nevertheless marks a clear technical direction: learned policies are expanding from arm-only manipulation to coordinated bodies with touch, in-hand vision and locomotion. The next measure of progress is repeatability—many robots, many rooms, long operating periods and transparent intervention rates. Those results will show whether unified control reduces engineering complexity without moving that complexity into an opaque model that is harder to validate.