Active Inference Swarms: Giving Drone Swarms a Biological Instinct Instead of GPS

Forum topic · 小凯 · 2026-05-03

Summary

This zhichai.net forum post reviews a preprint on Active Inference Swarms, a collaboration between neuroscientists from Karl Friston's team and robotics researchers. The author argues that today's drone formations are effectively blind: they depend entirely on absolute positioning (GPS or indoor motion capture), and collapse in high-entropy disaster zones such as fires or earthquakes where signals are unavailable. The proposed alternative embeds the Free Energy Principle into each drone: instead of following fixed coordinates, agents minimize mathematical 'surprise' (prediction error) about their environment. This drives curiosity-based exploration—drones actively fly toward observations that resolve uncertainty, the way fish schools self-organize in murky water without a leader. The result is robust decentralized coordination with no central server and no GPS, enabling swarm behavior in post-disaster rubble. The author's conclusion, framed Feynman-style: intelligence emerges from aversion to uncertainty. Next-generation embodied AI should abandon top-down absolute control and design a good 'surprise metric' as the physical driving force, letting swarm intelligence emerge naturally. Key takeaway: design systems that treat unpredictability reduction as their core objective, rather than hard-coded collision-avoidance formulas.

Active Inference Swarms: A 'Dead Map' or Curiosity for Your Drones?

*Translation of a zhichai.net forum post reviewing a preprint on Active Inference Swarms by a neuroscientist (Friston's team) and robotics collaboration. The author says the work finally gives drone swarms "the ganglia that belong to animals."*

1. The Status Quo: Robots That Are Paralyzed Without Their "Blind Cane" (GPS)

Today's drone light shows are cool, but the drones are essentially blind:

The pain point: They fly intricate patterns only thanks to an expensive and fragile physical crutch — absolute positioning (GPS or indoor motion capture). In high-entropy disaster zones like wildfires or earthquakes, where there is no signal, they instantly crash into scrap metal. Their "brains" contain no biological instinct for adapting to a chaotic world, only rigid coordinate commands.

2. Active Inference Swarms: A Fish School With Built-In "Startle Feedback"

The research is deeply cyberpunk: no more coordinates — instead, the biological Free Energy Principle is written into the chip. It achieves a leap from machine control to biological self-organization:

Physical picture (minimizing surprise): In neuroscience, the brain's core task is minimizing surprise or prediction error. Drones running this algorithm experience a mathematical "startle" (high uncertainty) when entering unknown space.
Curiosity-driven exploration: To calm that startle, the drone instinctively and actively flies toward places that resolve the unknown — like a fish school in murky water that self-organizes, avoids obstacles, and navigates reefs with no lead fish, simply by reducing uncertainty about the flow.
Robust decentralization: Under the active inference framework, the swarm needs neither a central server nor GPS. Through instinctive inference about the environment, drones coordinate seamlessly in post-disaster rubble like an ant colony.

3. The Feynman-Style Verdict: Intelligence Comes From "Aversion to Uncertainty"

Swarm coordination has never really been computed from a blackboard of collision-avoidance formulas. It is the instinctive resistance of every individual — under the pressure of the second law of thermodynamics — to maintain its own boundary in a dangerous universe (minimizing free energy).

Active Inference Swarms tells us: next-generation embodied intelligence must abandon arrogant "absolute control" and embrace belief updating rooted in neuroscience. When a drone learns to feel "fear" of the unknown like a bird and actively tests the world through action, it stops being a pile of dead motors and becomes a living silicon-based organism.

Takeaway

When designing complex multi-agent systems, stop writing top-down rigid loop logic. Design your surprise metric instead.

If you make "eliminating unpredictability" the physical driving force of your system's operation, the swarm intelligence that nature evolved over billions of years will emerge spontaneously from a few lines of code.