Insect-inspired drones get AI brains to race through tight spaces

Researchers at Shanghai Jiao Tong University have developed an innovative AI-based system that enables drone swarms to navigate complex, cluttered environments at high speeds without expensive hardware or human control. Unlike traditional modular drone navigation systems that separate tasks like mapping and obstacle detection—often leading to slow reactions and accumulated errors—the team created a compact, end-to-end neural network using differentiable physics. This approach allows the system to learn flight control directly through simulation and backpropagation, significantly improving learning speed and real-world performance. The drones rely on ultra-low-resolution 12x16 pixel depth cameras, inspired by insect compound eyes, to make real-time navigation decisions, achieving speeds up to 20 meters per second and a 90% success rate in cluttered spaces, outperforming previous methods. A key advantage of this system is its low cost and efficiency: the neural network runs on a $21 development board without requiring a graphics processing unit, making large-scale swarm deployment more accessible. The AI was trained entirely in simulation