Physics-Informed and Data-Driven Composite Disturbance Separation and Predictive Control for Airborne Electro-Optical Gimbals with Multiple Constraints

Yaokun Lu, Yifan Li, Zhixuan Zhang, Enmei Wang, Hao Teng, Jianzhong QiaoarXiv 2026 · arXiv

TL;DR

This paper introduces a hybrid control strategy for airborne gimbals that separates complex disturbances into periodic (vibration) and aperiodic (friction, gravity) components. By combining physics-informed frequency tracking with a "biomimetic" LSTM-based observer, the researchers achieved up to a 45.9% reduction in tracking errors compared to industry-standard methods while strictly respecting hardware and mission constraints.

What Problem Does This Paper Address?

Airborne electro-optical gimbals—the stabilized camera systems found on drones and surveillance aircraft—must maintain a rock-steady line-of-sight (LOS) while the vehicle is moving. This is exceptionally difficult because the system is bombarded by a "composite" of disturbances:

  1. Periodic Jitter: High-frequency vibrations from propellers and atmospheric turbulence.
  2. Aperiodic Torques: Slower-moving effects like gravity and gyroscopic forces, alongside fast, non-linear friction.
  3. System Constraints: Physical limits on motor torque, mechanical range, and the camera's field-of-view (FOV).
  4. Input Delay: The time lag between a command being sent and the motor reacting.

Most existing controllers treat these as a single "total disturbance." However, because these disturbances have vastly different physical origins and spectral characteristics, a one-size-fits-all estimation approach often leads to poor accuracy or control saturation.

Methods

The authors propose a Disturbance-Separated Model Predictive Control (MPC-DS) framework. The architecture is split into two specialized paths:

  • Adaptive Periodic Disturbance Observer (APDO): This is the "physics-informed" part. It uses a multi-harmonic model and a linear-regression frequency estimator to track vibrations in real-time. Crucially, these periodic predictions are fed directly into the MPC’s internal prediction model, allowing the controller to "see" and counteract future vibrations.
  • LSTM-based Disturbance Observer (LSTM-DO): This is the "data-driven" part, described as "biomimetic." A linear observer (the "brainstem") handles basic corrections, while a Long Short-Term Memory (LSTM) network (the "cerebellum") learns to predict complex aperiodic residuals and the specific changes that occur during the system's input delay.
  • Constraint Management: The MPC formulation uses "input-set tightening." It reserves a "headroom" margin for disturbances, ensuring that the commands sent to the motors never hit saturation limits, even when the environment gets rough.
Figure 1: Proposed MPC architecture with physics-informed and data-driven disturbance separation.

Datasets

The study did not use standard public datasets but instead relied on:

  • Identified Plant Models: Parameters for a commercial two-axis gimbal were identified using frequency sweeps.
  • Simulated Scenarios: Evaluation was performed on "Nominal" and "Calibrated" (more realistic, noisy) plant variants with injected disturbances (random walks, steps, and multi-harmonic vibrations).
  • Hardware Logs: The LSTM was trained offline using closed-loop data collected from the actual gimbal hardware to capture real-world friction and lag characteristics.

Key Insights and Ideas

  • Separation is Power: Treating periodic and aperiodic disturbances differently is physically motivated. Periodic signals are predictable over a horizon, whereas aperiodic signals are best compensated for at the current moment to avoid amplifying estimation errors within the optimizer.
  • Linear Regression vs. Gradient Descent: For frequency estimation, the authors found that a linear-regression approach is much faster at adapting to sudden changes in vibration frequency (e.g., when a drone changes motor speed) compared to traditional gradient-based laws.
  • Strategic Placement: A key "aha" moment was embedding periodic predictions inside the MPC optimizer while keeping aperiodic compensation outside. This hybrid placement maximizes accuracy while maintaining the stability of the optimization process.

Key Results

The proposed MPC-DS was tested against several benchmarks, including standard MPC and Linear Active Disturbance Rejection Control (LADRC):

  • Tracking Accuracy: In indoor hovering experiments on a multirotor UAV, the method achieved yaw and pitch errors (RMSE) of 0.770° and 0.350°, respectively.
  • Error Reduction: This represents a 25.1% (yaw) and 45.9% (pitch) improvement over LADRC.
  • Robustness: The system maintained high performance even with a 30% parameter mismatch in the underlying physical model, proving it can handle the wear-and-tear or manufacturing variances typical of real hardware.
  • Constraint Satisfaction: Unlike other predictive methods that frequently saturated the actuators (hitting 90%+ saturation in some tests), the MPC-DS stayed within limits 100% of the time.

Limitations and Future Work

While the results are impressive, the authors note a few areas for further exploration:

  • Flight Dynamics: The current experiments focused on indoor hovering. Future work needs to test the system during aggressive outdoor maneuvers where carrier motion is more violent.
  • LSTM Training: The data-driven component requires high-quality training logs. Developing ways to adapt the neural network online without compromising stability remains a challenge.
  • Computational Load: While the Intel N100 onboard computer handled the 20ms sampling period, scaling this to even higher-rate systems may require further optimization of the LSTM and QP (Quadratic Programming) solver.

Summary updated

Abstract

This paper studies the constrained pointing control problem of airborne electro-optical gimbals subjected to composite periodic and aperiodic disturbances. Under carrier vibration, input delay, and actuator constraints, achieving line-of-sight stabilization and tracking is challenging. A disturbance-separated model predictive control (MPC) scheme is proposed to compensate these heterogeneous disturbances according to their physical characteristics. An adaptive periodic observer predicts multi-harmonic vibrations over the MPC horizon using linear-regression frequency estimation. A linear observer estimates the aperiodic disturbance, and a bounded long short-term memory (LSTM) network predicts its change over the input delay. Embedding periodic predictions into the MPC model while applying aperiodic compensation outside the optimizer limits amplification of aperiodic estimation errors and improves pointing accuracy under frequency changes and parameter mismatch. On the stated headroom envelope, input-set tightening provides feasible commands within the prescribed limits, and a regional practical tracking bound is established under explicit input-sequence and predicted-state margin conditions. A UAV-mounted two-axis gimbal platform is assembled for comparative indoor hovering experiments, complemented by simulations under frequency changes and parameter mismatch. In the reported trials, the proposed method achieves yaw and pitch relative-angle root-mean-square errors (RMSEs) of 0.770 degrees and 0.350 degrees, reducing tracking errors by 25.1% and 45.9% compared with linear active disturbance rejection control.
BibTeX
@misc{lu2026physics,
  author = {Yaokun Lu and Yifan Li and Zhixuan Zhang and Enmei Wang and Hao Teng and Jianzhong Qiao},
  title = {{Physics-Informed and Data-Driven Composite Disturbance Separation and Predictive Control for Airborne Electro-Optical Gimbals with Multiple Constraints}},
  year = {2026},
  eprint = {2610.34849},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34849},
}
Figure