Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs
Making tiny flying robots smarter while using less power and memory
Researchers automated the process of shrinking and optimizing the artificial intelligence that controls nano-drones, cutting memory use in half and speeding up the drone's decision-making by 1.6 times. The optimized system let a Crazyflie nano-drone fly twice as fast as before, avoid obstacles more sharply, and navigate turns—all while using less than 2% of the drone's power budget.
Nano-drones could soon monitor crops, inspect infrastructure, or search buildings in disaster zones, but only if their onboard AI runs fast enough on battery-powered chips no bigger than a coin. This work removes the tedious hand-tuning that currently slows development, making it practical to deploy smarter autonomous drones at scale and letting researchers focus on new applications instead of wrestling with optimization details.