Figure Helix

2025-02-20 · Publisher : Figure AI · Domain : Robotics

Figure unveils Helix, a VLA controlling a humanoid's whole upper body: a 7-billion-parameter VLM "thinks slow" at 7-9 Hz and an 80-million-parameter policy acts at 200 Hz, fully onboard.

What changes

First VLA controlling a humanoid's entire upper body, fingers included, per Figure.
S2: a 7B VLM at 7-9 Hz; S1: an 80M policy at 200 Hz.
One set of weights for all tasks, no task-specific fine-tuning.
About 500 hours of teleoperation for training.
Runs entirely on embedded GPUs.

### Background On February 20, 2025, Figure AI, a Californian humanoid maker, unveiled Helix, its first in-house vision-language-action model, weeks after ending its partnership with OpenAI. Helix continuously controls the robot's wrists, torso, head and individual fingers from images and a natural-language instruction. ### Specifications - **System 2 (S2)**: a 7-billion-parameter open-weight VLM pretrained on internet data, understanding the scene at 7-9 Hz. - **System 1 (S1)**: an 80-million-parameter visuomotor policy turning S2's intent into continuous actions at 200 Hz. - **Data**: about 500 hours of multi-robot teleoperation, auto-labeled by a VLM. - **Deployment**: one set of weights for all tasks, running on low-power embedded GPUs. - **Demos**: two robots jointly put away groceries they had never seen, with the same network. ### Why it matters Helix makes the "think slow, act fast" architecture concrete on a full humanoid without offboard compute. NVIDIA adopted this two-system design with GR00T N1 a month later. ### Key takeaway Helix marks humanoid makers' entry into the proprietary foundation model race; Figure evolved it into Helix 02 and Helix 2.5.

Figure Helix

What changes

Links