Aggressiveness-Aware Learning-based Control of Quadrotor UAVs with Safety Guarantees

1. Problem Statement

Quadrotor UAVs operating in dynamic, cluttered environments must track aggressive trajectories with high precision. However, real-world flight is subject to complex, unmodeled dynamics (aerodynamic drag, wind gusts, rotor asymmetries) and external disturbances.

• The Core Conflict: Traditional geometric controllers on $SE(3)$ often rely on high feedback gains to reject these unknown disturbances and ensure tracking accuracy.
• The Consequence: High gains lead to feedback-induced aggressiveness, characterized by rapid, large variations in thrust and torque. This results in:
  • Increased energy consumption.
  • Actuator saturation.
  • Unsafe behavior near humans or obstacles.
  • Excitation of unmodeled high-frequency dynamics.
• The Gap: There is a lack of principled methods to explicitly trade off model accuracy, robustness, and control aggressiveness. Existing approaches often treat disturbance rejection and gain tuning as separate problems without quantifying the "aggressiveness" cost.

2. Methodology

The authors propose a Learning-Augmented Control Framework that integrates a learning-based oracle (specifically Gaussian Processes) with a nominal geometric controller to mitigate disturbances while minimizing feedback gains.

A. System Modeling

• Dynamics: The quadrotor is modeled on the Special Euclidean group $SE(3)$, separating known rigid-body dynamics from unknown generalized forces and moments ($f(x)$).
• Control Law: The control input $u$ is augmented as:
  $$u = h_{ff}(x, x_d) - K_{dyn}(x)\hat{f}(x) + h_{fb}(x)H e$$
  Where:
  • $h_{ff}$: Nominal feedforward term.
  • $\hat{f}(x)$: Learning-based oracle estimating the unknown disturbance.
  • $K_{dyn}$: Map distributing learned estimates to physical inputs.
  • $H$: Tunable feedback gain matrix.
  • $e$: Tracking error vector.

B. Defining "Aggressiveness"

The paper formalizes aggressiveness as a local input-sensitivity metric. It measures how strongly the commanded thrust and torques react to small variations in the tracking error:
$$s(H, x) = \left\| \frac{\partial}{\partial e} (h_{fb}(x)He) \right\| = \| h_{fb}(x)H \|$$
Higher values of $s(H, x)$ indicate more aggressive control behavior.

C. Learning-Based Oracle (Gaussian Processes)

• Oracle: A Gaussian Process (GP) is used to predict the unknown disturbance $f(x)$ based on a dataset of state-disturbance pairs collected during operation.
• Uncertainty Quantification: GPs provide not just a mean prediction but also a variance (uncertainty). The authors leverage this to compute high-probability error bounds ($\bar{\rho}_N$) on the model error over a compact set of trajectories.
• Safety Gating: To prevent unsafe corrections when the model is uncertain, a "gate" mechanism suppresses the learned compensation if the epistemic uncertainty exceeds a threshold.

D. Aggressiveness-Aware Gain Scheduling

The core innovation is casting the gain selection ($H$) as an optimization problem:

• Objective: Minimize feedback-induced aggressiveness ($s(H, x)$).
• Constraint: Ensure the tracking error converges to a prescribed practical exponential bound ($\|e(t)\| \leq \gamma_1 e^{-\gamma_2(t-t_0)}\|e(t_0)\| + \epsilon$).
• Mechanism: The required feedback gain is inversely proportional to the accuracy of the learning oracle.
  • High Model Accuracy (Low $\bar{\rho}_N$): The controller can use lower feedback gains to achieve the same tracking tolerance, reducing aggressiveness.
  • Low Model Accuracy (High $\bar{\rho}_N$): The controller must increase gains to maintain robustness, increasing aggressiveness.

3. Key Contributions

1. Formalization of Aggressiveness: Introduced a rigorous mathematical metric ($s(H, x)$) to quantify the sensitivity of control inputs to tracking errors, linking it directly to safety and actuator stress.
2. Learning-Augmented Stability: Proved that combining a nominal geometric controller with a learning-based disturbance estimator (with high-probability error bounds) guarantees practical exponential stability.
3. Principled Gain Scheduling: Developed a method to select feedback gains that minimize aggressiveness while satisfying tracking requirements, explicitly utilizing the probabilistic error bounds provided by the learning oracle.
4. Trade-off Analysis: Demonstrated theoretically and empirically that improving the disturbance model (via learning) allows for a reduction in feedback gains, thereby achieving safer, less aggressive maneuvers without sacrificing tracking performance.

4. Results

The framework was validated through numerical simulations under moderate and severe disturbance scenarios (DIST SCALE = 1 and 3).

• Performance vs. Aggressiveness Trade-off:
  • Fixed-Low Gains: Failed to meet tracking tolerance under severe disturbances ($\|e\| > 0.35$m).
  • Fixed-High Gains: Improved tracking but significantly increased transient aggressiveness (higher thrust/torque rates).
  • Aggressiveness-Aware (GP-Compensated): Achieved the tracking tolerance ($\|e\| \approx 0.028$m) with significantly lower aggressiveness than the fixed-high baseline.
• Quantitative Metrics:
  • Under severe disturbance, the learning-augmented controller reduced the transient thrust rate ($|\dot{T}|_{RMS}$) from ~10.4 N/s (fixed-high) to ~8.1 N/s, comparable to the low-gain baseline, while maintaining superior tracking accuracy.
  • The "gate" mechanism successfully suppressed compensation during high-uncertainty periods, ensuring safety.
• Online Learning: Experiments showed that updating the GP model online (residual learning) further refined performance, maintaining low aggressiveness while adapting to changing dynamics.

5. Significance

This paper addresses a critical bottleneck in autonomous UAV flight: the reliance on high-gain feedback for robustness. By integrating learning-based disturbance estimation with geometric control theory, the authors provide a principled pathway to safer, more energy-efficient flight.

• Safety: Reduces the risk of actuator saturation and unsafe maneuvers near obstacles.
• Efficiency: Lowers energy consumption by avoiding unnecessary aggressive control actions.
• Theoretical Rigor: Unlike "black-box" learning approaches, this method provides explicit probabilistic safety guarantees and stability proofs, making it suitable for deployment in safety-critical applications.
• Future Impact: The framework sets a standard for "aggressiveness-aware" control, suggesting that future UAV controllers should dynamically adjust their "stiffness" based on the confidence of their internal world models.