Recursive Gaussian Process compensation cuts quadrotor single-axis horizontal tracking error by 38% to 75%

It proposes using a Recursive Gaussian Process to compensate online for the nonlinear dynamics left uncanceled by a feedback linearization controller, together with a new Kalman-filter-based tuning algorithm designed for this purpose. Learning-based controllers have largely remained in simulation or offline implementations, constrained by UAV onboard computation; this work pairs the RGP's sparsity technique with Kalman-filter-based tuning so the compensation stage can run online. The text describes the RGP as a Bayesian model with a sparsity technique that reduces the algorithm's computational complexity, and states that the tuning algorithm was specifically designed for this purpose; the abstract does not give algorithmic details or complexity figures.

The entire RGP pipeline is deployed directly onboard a real quadrotor and validated experimentally for single-axis horizontal position control. Unlike prior learning-based controllers limited to simulation or offline use by onboard computation, this work brings online learned compensation onto a real flight platform. The platform is the ANT-X Lab drone, the validated task is single-axis horizontal position control, and the setting is indoor; the abstract reports no onboard compute load or real-time metrics.

With RGP added, tracking accuracy improves consistently across repetitive and non-repetitive sinusoidal references, with geometric mean root mean square error improvements of 38% to 75%. The comparison uses the same controller without RGP action as the baseline and spans multiple sinusoidal reference signals, indicating the gain is not confined to one trajectory type. The abstract reports geometric mean root mean square error improvements of 38% to 75% across all experiments and geometric mean tracking errors of 4.5 cm to 9.85 cm; per-experiment sample sizes and statistical tests are not given.

The result targets UAV control settings that need improved tracking accuracy under unmodeled disturbances or unmodeled dynamics, especially platforms with limited onboard computation that must run online. It shows that a sparsity-enabled Recursive Gaussian Process with Kalman-filter-based tuning can be realized on a real quadrotor in the form of single-axis horizontal position control, and provides an experimental starting point for extending similar online compensation to more degrees of freedom, outdoor environments, and stronger disturbances. For readers, the directly reusable elements are the structure of feedback linearization plus online Bayesian residual compensation, and the idea of designing the tuning algorithm and the compensation model together and putting both onboard.

The abstract does not state the degree of RGP sparsity, kernel choice, training data size, or online update frequency, nor does it report onboard compute load, real-time behavior, or failure cases. Experiments cover only indoor single-axis horizontal position control and sinusoidal-type references, without actually injecting the wind disturbances or aerodynamic effects mentioned as unmodeled factors, and without multi-axis or outdoor results. The text also notes that formal stability guarantees for learning-based controllers remain an open challenge, so the stability envelope of this scheme is not yet defined. Because the reading scope here is incomplete and figures and per-run data are missing, the judgments above about improvement magnitude and error levels rest only on the summary numbers in the abstract; the specific experimental conditions and number of repetitions remain open questions for the reader to check.