KNODE-DW MPC

Flying Quadrotors in Tight Formations using Learning-based Model Predictive Control

I tackled the challenge of enabling quadrotors to fly in tight formations and navigate in the presence of complex downwash airflows. To accomplish this, I combined first-principle physics models with neural ODEs to accurately model the dynamics of quadrotor downwash. While MPC can be employed to achieve exceptional control performance, it requires a precise model of the system dynamics. As such, I developed a learning-based MPC framework — KNODE-DW MPC — by integrating the hybrid physics and neural ODE model with MPC.

We obtained groundbreaking results with this framework such as a 40% performance improvement in trajectory tracking compared to nominal MPC in both simulations and physical experiments. I successfully flew two Crazyflie quadrotors in a stacked formation with only 12 cm of vertical separation—the first demonstration of its kind. This work resulted in a submission to ICRA 2025 where I am a co-first author.

Publication Link: Submitted to ICRA 2025

Publication PDF: PDF

Demo video at: Demo

KNODE-DW MPC: Top: A composite photo depicting two Crazyflie (CF) quadrotors flying in lemniscate trajectories, under a stacked formation with a commanded separation of 2 body lengths. This is achieved with our proposed framework. The frames used to create this photo are extracted from a video taken during a physical experiment. Bottom: Schematic of our framework.