Humanoids
Daily
Published on

Figure AI Reorganizes to Boost Humanoid Learning with 'Helix' AI Model

Authors
Two Figure humanoid robots collaborating
Figure's Helix AI enables its robots to understand and execute complex, long-horizon tasks like collaborative grocery organization, showcasing the model's potential in unstructured home environments. Image credit: Figure

Figure AI Reorganizes to Boost Humanoid Learning with 'Helix' AI Model

Figure AI is undergoing a significant internal restructuring to accelerate its artificial intelligence development. CEO Brett Adcock announced via social media today (May 29th) that the company executed its "largest re-org in Figure’s history," consolidating three separate teams into a unified AI group named Helix.

"Figure is an AI company at its core, and this consolidation will accelerate how quickly our robots learn and scale into market," Adcock stated, signaling a sharpened focus on the software driving its humanoid hardware.

Helix: A Generalist AI for Humanoids

The newly formed Helix group shares its name with Figure's latest AI model, detailed in a company blog post from February 2025. Helix is a Vision-Language-Action (VLA) model designed to provide generalist control for Figure's humanoid robots. The company claims Helix represents several breakthroughs:

  • Full Upper-Body Dexterity: Helix reportedly manages high-rate (200Hz) continuous control over the entire humanoid upper body, encompassing 35 degrees of freedom, including individual fingers, wrists, torso, and head movements. This allows for coordinated actions like head-tracking of hands while adjusting torso posture for optimal reach.
  • Multi-Robot Collaboration: Figure states Helix is the first VLA to operate simultaneously on two robots, enabling them to perform shared, long-horizon manipulation tasks with previously unseen items. For example, robots can collaborate on tasks like storing groceries based on natural language commands.
  • Versatile Grasping: Robots equipped with Helix can purportedly pick up a wide array of small household objects—even thousands of items not encountered during training—by following natural language prompts like "Pick up the [X]".
  • Unified Neural Network: A single set of neural network weights is used for all learned behaviors, including picking, placing, operating drawers, and inter-robot interaction, without requiring task-specific fine-tuning.
  • Onboard Processing: Helix is designed to run entirely on embedded, low-power-consumption GPUs on the robot, a step Figure describes as making it "immediately ready for commercial deployment."

The 'System 1, System 2' Approach

Figure explains that Helix employs a dual-system architecture, reminiscent of the 'thinking fast and slow' concept, to balance generalization with real-time responsiveness:

  • System 2 (S2): An internet-pretrained VLM (7B parameters) operating at 7-9 Hz. It handles scene understanding and language comprehension for high-level goals.
  • System 1 (S1): A faster (200 Hz) reactive visuomotor policy (80M parameters) that translates S2's semantic representations into precise, continuous robot actions.

This decoupled design allows S2 to perform complex reasoning while S1 executes and adjusts actions rapidly. The two systems are trained end-to-end.

Screenshot from Figure's website of a diagram explaining Helix
The dual-system architecture of Helix visualized. Image credit: Figure

Training and Efficiency

Figure highlights the data efficiency of Helix, stating it was trained on approximately 500 hours of teleoperated human demonstrations. This dataset, comparatively smaller than some other VLA training sets, was augmented using an auto-labeling VLM to generate hindsight instructions. The company emphasizes that despite this relatively modest data requirement, Helix can control a high-dimensional humanoid upper body and generalize to novel objects and tasks.

Advancing Humanoid Capabilities

The development of Helix aims to address the significant challenge of deploying robots in unstructured environments like homes, where an immense variety of objects and situations can be encountered. By enabling robots to understand natural language and generate new behaviors for unseen items on-demand, Figure hopes to overcome the limitations of traditional robot programming and imitation learning, which often require extensive manual effort or vast amounts of task-specific data. This focus on advanced AI is complemented by rapid hardware development; as highlighted in a recent announcement, the company's next-generation F.03 humanoid—successor to the Figure 02—is now walking, promising an even more capable platform for models like Helix.

Discuss on X
Subscribe to the newsletter