Neural Predictor for Flight Control with Payload

TL;DR

This work addresses the challenge of external disturbances from suspended payloads in tethered-UAVs by modeling the payload- and residual-induced force/torque as a dynamical system using a lifted-linear (Koopman-based) representation. The Neural Predictor combines a learned embedding with first-principles dynamics and is embedded in an MPC framework (NP-MPC) to improve real-time control, with a Lipschitz-constrained training regimen and a formal bound on prediction error. Empirical results show significant improvements in force/torque estimation accuracy and closed-loop tracking, including substantial gains over state-of-the-art estimators in simulations and real flights, and robustness to payload oscillations. The approach offers fast, interpretable, and data-efficient modeling for payload-carrying aerial robots, with practical impact on precise transport and manipulation tasks.

Abstract

Aerial robotics for transporting suspended payloads as the form of freely-floating manipulator are growing great interest in recent years. However, the force/torque caused by payload and residual dynamics will introduce unmodeled perturbations to the aerial robotics, which negatively affects the closed-loop performance. Different from estimation-like methods, this paper proposes Neural Predictor, a learning-based approach to model force/torque induced by payload and residual dynamics as a dynamical system. It yields a hybrid model that combines the first-principles dynamics with the learned dynamics. The hybrid model is then integrated into a MPC framework to improve closed-loop performance. Effectiveness of proposed framework is verified extensively in both numerical simulations and real-world flight experiments. The results indicate that our approach can capture force/torque caused by suspended payload and residual dynamics accurately, respond quickly to the changes of them and improve the closed-loop performance significantly. In particular, Neural Predictor outperforms a state-of-the-art learning-based estimator and has reduced the force and torque estimation errors by up to 66.15% and 33.33% while requiring less samples. The code of proposed Neural Predictor can be found at https://github.com/NPU-RCIR/Neural-Predictor.git.

Figures (9)

• Figure 1: The illustration of proposed framework: Neural Predictor cooperated with NMPC scheme.
• Figure 2: Force and torque estimation performance of Neural Predictor, NeuroBEM and NeuroMHE on a part of "Random points" test dataset (from 45s to 55s). The estimation errors on the whole "Random points" trajectory are $\text{RMSE}_{F}^{\text{NeuroMHE}}=0.260$, $\text{RMSE}_{\tau}^{\text{NeuroMHE}}=0.012$, $\text{RMSE}_{F}^{\text{NeuroBEM}}=0.530$, $\text{RMSE}_{\tau}^{\text{NeuroBEM}}=0.013$, $\text{RMSE}_{F}^{\text{NP}}\textbf{=0.088}$, $\text{RMSE}_{\tau}^{\text{NP}}=\textbf{0.008}$, where $F_{xy} = \sqrt{{F_x}^2+{F_y}^2}$, $\tau_{xy} = \sqrt{{\tau_x}^2+{\tau_y}^2}$, $F = \sqrt{{F_x}^2+{F_y}^2+{F_z}^2}$ and $\tau = \sqrt{{\tau_x}^2+{\tau_y}^2+{\tau_z}^2}$.
• Figure 3: The force and torque estimation RMSEs of Neural Predictor that is trained with different sample size on "Race track_1" test dataset. The gray and blue dash-dotted lines denote force and torque estimation RMSEs of NeuroMHE which is trained with 4K samples on "Race track_1" test dataset, respectively.
• Figure 4: The schematic of real-world flight experiments setup.
• Figure 5: Estimation of the external force in Z-axis when additional payload is attached.

Theorems & Definitions (4)

• Theorem 1
• proof
• Lemma 1
• proof