Helix 02
Figure AI · USA · 2026
Figure's three-tier robot brain — the first to put walking, balance and manipulation under one learned policy.
Helix 02, announced by Figure on 27 January 2026, extends the company's vision-language-action stack from the upper body to the entire machine. Where the original Helix split the brain in two, Helix 02 adds a third layer underneath. System 2 still runs on a slow timescale, interpreting scenes, understanding language and sequencing behaviours. System 1 runs at 200 Hz and translates perception into full-body joint targets. Underneath them, System 0 executes at 1 kHz and handles balance, contact and coordination across the whole body. System 0 is the interesting part, and not because of its size — it is a 10-million-parameter network, tiny by the standards of this catalogue. It is interesting because of what it replaced: 109,504 lines of hand-engineered C++ control code, swapped for a learned whole-body motion prior. It was trained on over 1,000 hours of joint-level retargeted human motion data, entirely in simulation, across more than 200,000 parallel environments with extensive domain randomisation. That is the mechanism by which a humanoid that previously learned to manipulate can now also learn to walk, carry and recover without a separate hand-written controller for each. The result is a single pixels-to-whole-body network: head cameras, palm cameras, fingertip tactile sensors able to detect forces as small as three grams, and full-body proprioception feed straight through to joint-level control of legs, torso, head, arms, wrists and individual fingers. Figure's flagship demonstration was a four-minute, end-to-end autonomous task in which a robot walks to a dishwasher, unloads it, crosses a room, stacks items in cabinets, then loads and starts the machine — 61 loco-manipulation actions in the correct order, with implicit error recovery and no reset. Two later posts on Figure's news page, Helix 02 Living Room Tidy (9 March 2026) and Helix-02 Bedroom Tidy (8 May 2026), continue the same line of household autonomy work. One honest gap: Figure's announcement does not restate parameter counts for System 2 and System 1, does not disclose the onboard hardware for this version, and says nothing about weight availability. The 7B/80M figures in this catalogue belong to the original Helix and should not be assumed to carry over unchanged. The original profile remains in the catalogue as the entry that established the architecture. Where the next round of training data comes from was answered on 25 August 2026, when Figure came out of stealth with Index — a consumer app that pays people to film themselves doing ordinary chores. The company reports over 16 million uploaded videos from more than 44,000 weekly active contributors in 108 countries, 15 million dollars paid out, and a commitment to spend over one billion dollars on data and compute in the following 12 months. System 0 learned whole-body motion from retargeted human motion in simulation; Index is the pipeline built to feed the layers above it with real-world human video.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!