Abstract
Traditional Nonlinear Model Predictive Control (NMPC) relies on manually tuned, fixed cost weights that often fail to generalize across diverse robotic maneuvers. This paper presents a novel NMPC framework that automatically adapts its cost weights in real-time. By formulating weight selection as a min-max optimization problem and solving it via a two-stage alternating Real-Time Iteration (RTI) algorithm, the authors achieve up to a 70% improvement in trajectory tracking accuracy for quadrotors.
Introduction
Nonlinear Model Predictive Control (NMPC) is a cornerstone of modern robotics, particularly for agile systems like quadrotors. NMPC functions by solving an optimization problem to determine a sequence of control actions that minimize a cost function—typically a weighted sum of state error and control effort. However, determining the optimal weights ($Q$ and $R$ matrices) is notoriously difficult, often requiring extensive manual tuning by experts. A single set of weights rarely suffices for both slow hovering and aggressive high-speed flight. Furthermore, static weights do not inherently prioritize control smoothness, which is critical for hardware longevity.
Methodology: Online Weight Adaptation
The core contribution of this work is an Online Weight-adaptive NMPC framework. Instead of treating the weight matrix $Q$ as a static hyperparameter, the authors incorporate it into the optimization objective. To maintain the real-time performance necessary for flight control, they propose a two-stage alternating Real-Time Iteration (RTI) algorithm.
The algorithm operates in two distinct phases:
- Stage 1 (State and Control Prediction): The weights are held constant, and a standard Quadratic Program (QP) is solved to predict the optimal state trajectory and control sequence.
- Stage 2 (Weights Update): The predicted trajectory is fixed, and the weight matrix is updated based on the predicted error. If a high error is predicted for a specific state, its corresponding weight is increased to prioritize its correction in the next iteration.
The authors derived a closed-form solution for this update using an exponential function of the error, ensuring computational efficiency. This approach is theoretically grounded in Metric Learning, where the controller effectively learns a Mahalanobis distance metric to measure its own performance dynamically.
Experimental Setup
The framework was evaluated in high-fidelity simulations using a quadrotor model. Four distinct reference trajectories were tested:
- Aggressive T1 & T2: High-speed paths with sharp turns.
- Circle T3: A smooth, continuous circular path.
- Diamond T4: A trajectory characterized by abrupt direction changes at corners.
The performance was benchmarked against the standard NMPC implementation from the ACADO toolkit.
Key Results and Discussion
The adaptive controller demonstrated significant performance gains over traditional fixed-weight methods.
Tracking Accuracy
In aggressive maneuvers (T1 and T2), the weight-adaptive controller reduced positional error by up to 70%. Notably, the system achieved high accuracy even with short prediction horizons ($N=8$), where standard NMPC typically struggles.
The $\lambda$ Parameter and Smoothness
The authors introduced a single parameter, $\lambda$, to control the sensitivity of the weight updates. This simplifies the tuning process from adjusting dozens of matrix entries to a single intuitive dial. The results highlighted a trade-off between accuracy and control smoothness (measured by Total Variation, TV). While low $\lambda$ values (e.g., 0.01) maximize accuracy, moderate values (e.g., 0.67) provide a "sweet spot" that balances high precision with smooth control commands suitable for real-world hardware.
Robustness
The system maintained its performance advantage even in the presence of simulated sensor noise, indicating that the adaptive mechanism is robust to real-world uncertainties.
Conclusion and Future Work
This research demonstrates that moving from static to adaptive cost weights significantly enhances NMPC performance for complex robotic tasks. While currently validated in simulation, the computational efficiency of the closed-form weight update suggests high feasibility for real-time hardware deployment. Future work will focus on automating the selection of the $\lambda$ parameter and conducting physical flight tests to account for real-world dynamics like motor lag and environmental disturbances.
Summary updated