Figure AIUnited States2025
Helix
An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.
VersionsHelix 2025-02-20Helix 02 2026-01-27Helix 2.5 2026-09-17
The developer has released this model in several versions. Each figure applies to the version its source names.
Key figures as stated by the developer
4 of 6 stated · 2 gaps-
Model type B
vision-language-action -
Controls A
arms and handswhole bodywalking -
Inputs A
cameralanguagerobot statetouch - Access not stated
- Parameters not stated
- Base model E no
How to read the figures
- 14kgA30.9lbA
- The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
- 74.5lb
- Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.
Versions and notes in the key figures (6)
Model type
Developer wording: "Helix: A Vision-Language-Action Model for Generalist Humanoid Control".
Controls
Helix 2025-02-20: "S1 outputs full upper body humanoid control at 200hz, including desired wrist poses, finger flexion and abduction control, and torso and head orientation targets" (upper body only; torso and head targets have no value of their own in the closed list). Helix 02 2026-01-27: "Complete joint-level control of the entire robot - legs, torso, head, arms, wrists, and individual fingers.". Project Go-Big (18 September 2025, version not named): "A single Helix neural network now outputs both manipulation and navigation". Helix 2.5 2026-09-17: "Zero-shot whole-body autonomy across 30 real homes", "spanning locomotion, rigid and deformable manipulation, bimanual coordination, and active perception".
Versions
- Helix 2025-02-20
- arms and hands
- Helix 02 2026-01-27
- arms and hands, whole body, walking
- Helix 2.5 2026-09-17
- arms and hands, whole body, walking
Inputs
Helix 2025-02-20: "monocular robot images and robot state information" plus "natural language commands specifying desired behaviors". Helix 02 2026-01-27: "Head cameras, palm cameras, fingertip tactile sensors, and full-body proprioception."; language through System 2, which "reasons slowly about goals: interpreting scenes, understanding language, and sequencing behaviors". Helix 2.5 2026-09-17: not stated.
Versions
- Helix 2025-02-20
- camera, language, robot state
- Helix 02 2026-01-27
- camera, language, robot state, touch
Access
Note Figure calls Helix "Figure's proprietary vision-language-action AI".
Parameters
Note Figure gives parameters per layer only, never one total. Helix 2025-02-20: "S1, an 80M parameter cross-attention encoder-decoder transformer"; "S2 is built on a 7B-parameter open-source, open-weight VLM". Helix 02 2026-01-27: System 0 is "A 10M-parameter neural network that takes full-body joint state and base motion as input"; no count for System 1 or 2. Helix 2.5 2026-09-17: not stated.
Base model
Note Helix 2.5 2026-09-17: "Helix 2.5 was itself pretrained from random initialization entirely on Index, unlike Helix 02, which started from a pretrained vision-language model." Helix 2025-02-20: "S2 is built on a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data" (model not named). Helix 02 2026-01-27: a pretrained vision-language model (not named).
Versions
- Helix 2025-02-20
- open-weight 7B VLM (not named)
- Helix 02 2026-01-27
- pretrained VLM (not named)
- Helix 2.5 2026-09-17
- none (trained from scratch on Index)
Full specifications
9 of 12 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Input and output | ||
| Model type | vision-language-action | B Developer wording: "Helix: A Vision-Language-Action Model for Generalist Humanoid Control". |
| Inputs | camera language robot state touch | A Helix 2025-02-20: "monocular robot images and robot state information" plus "natural language commands specifying desired behaviors". Helix 02 2026-01-27: "Head cameras, palm cameras, fingertip tactile sensors, and full-body proprioception."; language through System 2, which "reasons slowly about goals: interpreting scenes, understanding language, and sequencing behaviors". Helix 2.5 2026-09-17: not stated. Versions
|
| Controls | arms and hands whole body walking | A Helix 2025-02-20: "S1 outputs full upper body humanoid control at 200hz, including desired wrist poses, finger flexion and abduction control, and torso and head orientation targets" (upper body only; torso and head targets have no value of their own in the closed list). Helix 02 2026-01-27: "Complete joint-level control of the entire robot - legs, torso, head, arms, wrists, and individual fingers.". Project Go-Big (18 September 2025, version not named): "A single Helix neural network now outputs both manipulation and navigation". Helix 2.5 2026-09-17: "Zero-shot whole-body autonomy across 30 real homes", "spanning locomotion, rigid and deformable manipulation, bimanual coordination, and active perception". Versions
|
| Degrees of freedom controlled | 35DoF | B Stated for Helix 2025-02-20 only, for the humanoid upper body: "Helix coordinates a 35-DoF action space at 200Hz" ("high-rate, dexterous control of the entire humanoid upper body"). Helix 02 2026-01-27 and Helix 2.5 2026-09-17 control the whole body and give no number. Versions
|
| Size and speed | ||
| Parameters | not stated | Note Figure gives parameters per layer only, never one total. Helix 2025-02-20: "S1, an 80M parameter cross-attention encoder-decoder transformer"; "S2 is built on a 7B-parameter open-source, open-weight VLM". Helix 02 2026-01-27: System 0 is "A 10M-parameter neural network that takes full-body joint state and base motion as input"; no count for System 1 or 2. Helix 2.5 2026-09-17: not stated. B |
| Control rate | 1,000Hz | A The fastest stated layer. Helix 02 2026-01-27 adds System 0 "with a new foundation layer: System 0"; it "executes at 1 kHz, handling balance, contact, and coordination across the entire body"; System 1 runs at 200 Hz. Helix 2025-02-20: "S1 outputs full upper body humanoid control at 200hz"; System 2 runs at 7-9 Hz. Helix 2.5 2026-09-17: not stated. Versions
|
| Where it runs and what it is built on | ||
| Runs where | on the robot | C "Helix controls the full loop: perception, movement, and reasoning, on board and in real time." Helix 2025-02-20: "Helix is the first VLA that runs entirely onboard embedded low-power-consumption GPUs". |
| Base model | no | Note Helix 2.5 2026-09-17: "Helix 2.5 was itself pretrained from random initialization entirely on Index, unlike Helix 02, which started from a pretrained vision-language model." Helix 2025-02-20: "S2 is built on a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data" (model not named). Helix 02 2026-01-27: a pretrained vision-language model (not named). E Versions
|
| Openness | ||
| Access | not stated | Note Figure calls Helix "Figure's proprietary vision-language-action AI". D |
| License | Proprietary | D Figure's word: "Figure's proprietary vision-language-action AI". |
| The developer’s own claims | ||
| Learns new tasks | needs retraining | E Developer claim. Helix 2.5 2026-09-17: "tasks were specified through fine-tuning data collected elsewhere"; "matched its success rate while using half as much adaptation data". Helix 2025-02-20: "without any task-specific demonstrations or extensive manual programming"; "new skills can be specified on the fly with language". Helix 02 2026-01-27: not stated. Versions
|
| Works across robots | not stated | Note Figure's claims concern several units of its own robot: "we were able to apply the same policy, initially trained on a single robot's data, to multiple additional robots", after correcting for "small individual robot hardware variations"; "both robots operate using identical Helix model weights, eliminating the need for robot-specific training or explicit role assignments". F Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count. |
Notes
- Figure describes Helix as a generalist humanoid Vision-Language-Action model that learns and improves over time as it acquires new skills.
- Helix 02 (27 January 2026) adds System 0, a whole-body controller that executes at 1 kHz below System 1 and System 2.
- Helix 2.5 (17 September 2026) was pretrained from random initialization on Index, Figure's dataset of human behaviour, not from a vision-language model.
Sources
-
A
Manufacturer product page. www.
figure. · Retrievedai/ news/ helix-02 -
B
www.
figure. · Retrievedai/ news/ helix -
C
www.
figure. · Retrievedai/ helix -
D
www.
figure. · Retrievedai/ news/ introducing-figure-03 -
E
www.
figure. · Retrievedai/ news/ helix-2-5-zero-shot-30-home-generalization -
F
www.
figure. · Retrievedai/ news/ helix-logistics
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.