Illustration brief the model wrote
A humanoid robot at a kitchen counter, loading plates into a dishwasher. Its hands have textured tactile grips and small palm cameras. The robot walks smoothly while holding a plate, one foot stepping forward. Dishwasher door open to right, plates stacked on counter left. No faces or logos.
Introducing Helix 02: Full-Body Autonomy
Helix 02 replaces hand-coded full-body control with a learned neural system for continuous multi-minute real-world tasks like loading a dishwasher.
What changed
ModelBefore, full-body control for humanoid robots required 109,504 lines of hand-engineered C++ code for balancing, walking, and manipulation. This paper introduces a learned System 0 controller trained on over 1,000 hours of human motion, enabling continuous 4-minute, 61-action tasks without resets, and adds tactile and palm-camera dexterity for fine manipulation such as single-pill extraction and syringe dosing.
So what, in plain English
ModelThink of a robot that used to need a separate playbook for every muscle—now it learns a dance by watching human videos, so it can walk, balance, and handle objects smoothly at the same time. This means it can do real chores like loading a dishwasher from start to finish without stopping and without someone writing instructions for every step.
Contribution to industry
ModelThis paper replaced massive hand-engineered C++ code with a learned neural prior for whole-body control, dramatically reducing engineering effort for deployment. It also demonstrated tactile sensing and palm cameras for precision manipulation (e.g., pill extraction, syringe dosing) on a real-world 4-minute dishwasher task. Figure AI likely uses this to deploy its humanoid robots in home or warehouse settings, and the 1,000+ hour human motion dataset provides a training resource for the field.
Typed judgments
Jev★★★★☆ robots · Physical AI 97%
Real hardware 79% · Open code / data 8% · Foundation-model based 28% · Survey / benchmark-only 2%
Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.
Editor's notes
Summary. Extends Helix to the whole body with a learned System 0 controller trained on over 1,000 hours of human motion, completing a continuous 4-minute, 61-action dishwasher task with walking, manipulation and balance and no resets.
Industry. Replaced 109,504 lines of hand-engineered whole-body C++ control with a neural prior and added tactile and palm-camera dexterity such as single-pill extraction and 5 ml syringe dosing.