SwarmMind: Bio-Inspired Collective Learning for Resilient Drone Swarms

Project Description

In nature, no individual learns in isolation. When a bird in a flock encounters a toxic insect or a predator, the entire group adapts — rapidly and without a central coordinator. Our project brings this same principle to autonomous drone swarms.SwarmMind introduces a bio-inspired social learning framework that enables groups of drones to collectively improve their navigation strategies in real time, directly from deployment experience. The core idea is elegantly simple: when a drone crashes or makes a critical error, rather than that failure being wasted, the drone uses its remaining onboard resources to analyze the incident and retrain its navigation model on the spot. This updated model — the lesson extracted from the crash — is then shared with neighboring drones via local wireless communication. The swarm learns together, adapts together, and avoids repeating the same deadly mistake.This approach follows the paradigm of Decentralized Learning and Execution (DLE): there is no central server, no global supervisor, and no pre-configured environmental map. Each drone acts locally, learns locally, and shares knowledge locally. Over time, the swarm builds a shared collective memory, analogous to cultural knowledge accumulation in animal communities.This project is carried out at the Center for Project-Based Learning, ETH Zürich, and Politecnico di Torino, Italy, leveraging international expertise in embedded systems, low-power communication, and onboard machine learning. Experiments conducted in both simulated and real-world hazardous environments demonstrate that socially-enabled swarms dramatically reduce repeated failures and adapt far faster than swarms with static or individually trained controllers.Are you interested? Contact us!

 Join Us to Work On

• Failure-Driven Onboard Learning — Designing lightweight neural network retraining pipelines that run on resource-constrained drone hardware after a crash event
• Peer-to-Peer Model Sharing — Developing efficient inter-drone communication protocols to propagate learned updates across the swarm in real time
• Swarm Navigation in Hazardous Environments — Testing and validating collective adaptation in cluttered, dynamic, or GPS-denied scenarios
• Bio-Inspired Swarm Algorithms — Translating biological collective intelligence mechanisms into scalable robotics frameworks
• Novel Sensors: Learn on how novel and lightweight sensors techniques can improve robotic perception and resilience to unknow environments.

Contacts       

• Tommaso Polonelli, 
• Daniele  Jahier Pagliari, 
• Alessio  Burrello,