Meta-learning MPC improves quadrotor tracking in controlled strong-wind flight tests
Researchers tested a meta-learning augmented model-predictive controller that estimates residual aerodynamic force online and feeds it back into the MPC model. In their fan-array experiments, with nominal wind settings up to 16 m/s measured 1 m downstream of the fans, the method improved average tracking performance by 50.6% over a GP-MPC baseline in real flight and 73.5% in simulation; the maximum reported reductions were 59.7% and 86.5%, respectively.
Fri, 2 Oct 2026
Wind compensation improves when the controller adapts a learned disturbance model online rather than relying only on a fixed aerodynamic model. That can be useful for inspection or mapping trajectories where cross-track error matters.
The wind field was generated in a controlled experimental setup, so the percentages should not be treated as a guarantee in gusty outdoor turbulence. Reproduce the test with the intended airframe, payload, wind spectrum and compute budget before using the result to expand an operating envelope.