top of page
Buscar

FIGURE AI’S HELIX MODEL

  • Foto do escritor: Adam
    Adam
  • 2 de set.
  • 2 min de leitura

In February 2025, Figure AI unveiled Helix, a groundbreaking Vision-Language-Action (VLA) model designed for versatile humanoid control. It unifies natural perception, language understanding, and continuous motor control into a single neural system.

  • Full upper-body control: Helix delivers precise, real-time commands to a humanoid’s torso, head, wrists, and individual fingers.

  • Collaborative multi-robot capability: It is the first VLA that can simultaneously operate two robots, coordinating long-horizon manipulation tasks with objects never seen before.

  • Unified neural approach: One single set of weights handles picking, placing, drawer operation, coordination, and more.

Dual-System Architecture: Fast and Smart

Helix leverages a two-part cognitive structure reminiscent of human thinking:

  • System 2 (S2): A 7B-parameter Vision-Language Model that models scene understanding and interprets natural language—operating at 7–9 Hz.

  • System 1 (S1): An 80M-parameter transformer-based visuomotor policy that executes rapid, continuous control at 200 Hz. This separation allows Helix to plan thoughtfully but act swiftly.

Trained for the Real World

Helix has been trained on around 500 hours of teleoperated robot behavior, paired with automatic language labeling—allowing it to understand contextual intent without any object-specific training.

Why Helix Matters

  1. Instant Generalization: Helix can pick up novel household items based on everyday language prompts—no prior exposure needed.

  2. Commercial Viability: It runs completely on-board embedded low-power GPUs—ready for real-world deployment, not just research labs.

  3. Scalable Robotics: This model marks a key shift toward robots that can learn and adapt on demand, using language as the interface—not manual coding or thousands of demonstrations.

Key Takeaways

Feature

Highlight

Helix Launch

First generalist VLA for humanoid upper-body control

Two-System Design

S2 for reasoning (7–9 Hz), S1 for action (200 Hz)

Unified Learning

Single neural network handles diverse tasks

Generalization

Works with novel objects upon natural language prompts

Deployment Ready

Runs on embedded hardware—no cloud needed

Helix signals a seismic shift in robotics — not just as a tool, but as a cognitive companion that bridges human intent with precise physical execution.

 
 
 

Comentários


Tel: +351 913 887 421

Rua de Diu 414
4150-272
Porto

Portugal

SIGN OUR NEWSLETTER

DISCOVER OUR LATEST PROJECTS AND INNOVATIONS

Email sent successfully!

© 2025 BY PIXELGRAMMAR LDA - All rights reserved.

  • LinkedIn
bottom of page