From Kahneman to Figure.ai
- Tommaso Pardi
- Mar 3
- 3 min read
Why Bother Giving Robots Two Brains?
In the fast-paced world of robotics, control systems are evolving rapidly. Companies like Figure.ai are pushing the limits, and their latest approach, Helix, is one of the most intriguing yet. Instead of relying on a single AI model to handle everything, Helix borrows a page from human cognition, using a two-system approach.
If that sounds familiar, it’s because it closely resembles Daniel Kahneman’s System 1 and System 2 theory—the idea that we humans have fast, intuitive thinking and slow, deliberate reasoning working together. Now, robots are getting their own version of this cognitive division.
So, what happens when a humanoid robot gets a dual-brain upgrade? Let’s break it down.
The Helix Diaries: Teaching Robots to Think Fast and Slow
Most robotic control systems fall into one of two categories: reactive and fast, or deliberate and intelligent. The problem? Traditional robots are usually one or the other—they either act quickly but dumbly or think carefully but move sluggishly. Helix solves this by combining both.
Helix is a Vision-Language-Action (VLA) model, meaning it lets robots see, understand commands, and act fluidly in real-time. Unlike older systems that require predefined scripts, Helix enables robots to adapt on the fly. In a demo, Figure’s humanoid robots worked together to put away groceries—one robot held the fridge door while the other placed items inside. No pre-programmed steps, just real-time problem-solving.
But how does Helix pull this off? By splitting intelligence into two layers:
System 2 (“The Thinker”): A large vision-language model (~7B parameters) that processes the environment, interprets language, and plans actions. Smart, but slow (~7 Hz).
System 1 (“The Doer”): A smaller motor control model (~80M parameters) that executes movements in real time at 200 Hz. Fast, but not “deep.”
The key? System 2 thinks ahead, and System 1 handles the real-time execution.
Peeling Back the Layers (Literally)
Once I started diving into Helix’s structure, the similarities to human cognition became crystal clear:
System 1 (Fast Reflexes) → Controls movement in real-time, ensuring the robot can respond to physical interactions on the fly.
System 2 (Deep Thinking) → Interprets the world, understands high-level instructions, and passes decisions to System 1.
By separating these two functions, Helix achieves both real-time responsiveness and adaptable intelligence—something that’s been a bottleneck in humanoid robotics for years.
Why This Matters
Humanoid robots aren’t just research toys anymore—they’re being developed for real-world use. Helix’s two-system approach unlocks huge advantages:
✅ Generalization – Robots can learn new skills instantly without requiring thousands of training samples.
✅ Scalability – One AI model controls multiple robots, allowing for collaboration (like passing objects between bots).
✅ Real-World Adaptation – Robots can react naturally to their environment instead of following rigid programming.
By applying Kahneman’s cognitive theory to robotics, Figure.ai has created a humanoid AI that thinks like us—fast when needed, careful when required.
Final Thoughts
Pre-built AI models are great. Understanding them is better. If you ever find yourself just fine-tuning models instead of truly engineering them, step back and dig deeper.
Helix is a glimpse into the future—one where robots can think, react, and adapt in human-like ways. Maybe, just maybe, we’re finally on the verge of humanoids that truly “get” us.
Now, onto the next challenge—maybe implementing System 1 & 2 myself? 🤔
Comentários