Curiosity · AI Model
Figure Helix (Figure 02)
Helix is Figure AI's generalist brain for the Figure 02 humanoid robot. It uses a dual-system architecture: System 2 is a 7B-class vision-language model that reasons at ~7–9 Hz about scene, goals, and language; System 1 is an 80M-parameter transformer policy that takes S2's latent as a conditioning vector and outputs 200 Hz whole-upper-body control (arms, torso, head, fingers). One network controls two robots collaboratively, zero-shot picks up household items it has never seen, and follows free-form natural-language commands — a milestone demo for commercial humanoids.
Model specs
- Vendor
- Figure AI
- Family
- Helix
- Released
- 2025-02
- Context window
- 1 tokens
- Modalities
- text, vision, code
Strengths
- Dual-system decouples slow reasoning from high-rate control
- 200 Hz fine-grained finger control — not just pick-and-place
- Single network generalises across multiple robots and tasks
- Proven on commercial-grade humanoid hardware
Limitations
- Proprietary — not released publicly, no weights or API
- Limited independent third-party evaluation
- Tied tightly to the Figure 02 hardware platform
- Requires Figure-scale data collection pipeline to replicate
Use cases
- Humanoid manipulation in domestic and warehouse settings
- Collaborative dual-robot tasks from shared instruction
- Voice-driven household chores
- Whole-upper-body coordinated control research
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| Zero-shot household-object grasping (Figure demo) | qualitative success on thousands of novel items | 2025-02 |
| Multi-robot collaborative task (two Figure 02) | first public demo | 2025-02 |
Frequently asked questions
What is Helix?
Helix is Figure AI's generalist vision-language-action model that runs on the Figure 02 humanoid. It combines a slow VLM planner with a fast 200 Hz visuomotor policy for whole-upper-body control.
Is Helix available to developers?
No — Helix is proprietary to Figure AI and currently deployed only on its own humanoid hardware.
How does Helix compare to π0 and RT-2?
Helix targets full humanoid upper-body control at 200 Hz and dual-robot collaboration, whereas π0 focuses on dexterous manipulation across varied embodiments and RT-2 emphasises semantic generalisation from web-scale co-training.
Sources
- Figure AI — Introducing Helix — accessed 2026-04-20
- IEEE Spectrum — Figure 02 Helix coverage — accessed 2026-04-20