These systems convincingly demonstrate the latency and dynamic-range advantages of events, but they execute their networks off-board or on onboard computers drawing typically 10 W or more, an order of magnitude beyond the power and payload budget of a nano-drone. At the extreme low-power end, a fully neuromorphic vision-to-control loop on a flying drone used spiking networks on the Intel Loihi research chip for optical-flow-based flight, landing, and hovering. Goal-directed navigation through cluttered environments, computed fully on board a commercially available microcontroller, has remained open—this is the capability demonstrated here.
Event-Based Drone Flight
Real-World Deployment
The N6 companion board and the Crazyflie exchange messages over a two-wire UART link at 115,200 baud. The N6 runs only the policy, while the Crazyflie’s STM32F405 runs the low-level stabilization. Framing carries a sync header and checksum, so corrupted packets are discarded rather than acted on.
Acknowledgements
The authors would like to thank Chaoxiang Ye, Yilun Wu, Hang Yu, Shenqi Wang, Rafael Pérez, and Elijah Ang for their insights during this project.
Frequently Asked Questions
What advantage do event-based systems demonstrate? They demonstrate latency and dynamic-range advantages.
What power budget challenge do nano-drones face? Onboard computers drawing 10 W or more exceed a nano-drone’s power and payload budget by an order of magnitude.
What does the N6 companion board run? The N6 runs only the policy.
How does the Crazyflie handle low-level stabilization? Its STM32F405 runs the low-level stabilization.