TL;DR
This paper bridges the gap between robust control theory and real-world robotics by implementing the first robust, perception-based controller on a physical quadrotor. By introducing a novel "imitation cost" function within the System Level Synthesis (SLS) framework, the authors provide a practical way to take a well-tuned, standard controller and "robustify" it against the inevitable drift and noise of visual-inertial odometry (VIO).
What Problem Does This Paper Address?
In robotics, controllers and state estimators are usually designed independently—a concept known as the separation principle. Controllers often assume they are receiving perfect state estimates. However, for lightweight robots operating in GPS-denied environments, state estimation relies heavily on Visual-Inertial Odometry (VIO). VIO is prone to drift over time, and its accuracy is highly dependent on the robot's movements and environmental lighting.
When perception degrades, standard controllers can overreact to noisy estimates, causing the system to misbehave or crash. While robust control theory offers mathematical guarantees to bound these errors, applying it to complex, real-world hardware is notoriously difficult. The authors tackle this gap, seeking to design a controller that explicitly accounts for perception error, guarantees bounded responses to that error, and is actually feasible to tune and deploy on a real, SWaP (Size, Weight, and Power) constrained quadrotor.
Methods
The authors model the quadrotor as a Linear Time-Invariant (LTI) system (linearized around hover) and use System Level Synthesis (SLS) to parameterize the achievable closed-loop behaviors of the system.
Their approach involves three main steps:
- Data-Driven Error Modeling: They fly the quadrotor to collect ground-truth vs. VIO data, characterizing the perception error using an $S$-slope. The $S$-slope acts as a worst-case bound on how fast the perception error degrades as the drone moves away from states it saw during training.
- Robustness Constraint: They inject this $S$-slope into the SLS optimization problem as a hard constraint, ensuring the system's response to perception errors remains mathematically bounded.
- The Imitation Cost Function (Novelty): Traditional robust controllers (like LQG or $L_1$) are incredibly difficult to tune using standard $Q$ and $R$ weight matrices. To solve this, the authors introduce an imitation cost. Instead of building a behavior from scratch, the optimization minimizes the difference between the robust controller and a pre-existing "nominal" controller (like a well-tuned PD controller), subject to the robustness constraint.
Datasets
No standard public datasets were used. Instead, the authors generated custom training and evaluation data:
- Simulation: Data collected using ROS/Gazebo with a simulated Multi-state Constraint Kalman Filter (MSCKF-VIO).
- Hardware: Data collected using a 256g Snapdragon Flight quadrotor equipped with a downward-facing fisheye camera. Ground truth was provided by a VICON motion capture system.
Key Insights and Ideas
- Theory vs. Reality of the $S$-slope: The strict mathematical $S$-slope of the real-world perception error was too conservative. A few noisy outlier measurements caused the theoretical bound to spike, making the robust optimization problem infeasible. The authors had to manually relax this bound to synthesize a working controller, highlighting a friction point between pure robust control theory and messy hardware.
- The Power of Imitation: The imitation cost function is a massive quality-of-life improvement for control engineers. It allows practitioners to keep the intuitive tuning of a PD controller while wrapping it in the safety guarantees of SLS.
- The "Softness" of Robust Control: The robust controller naturally learns to be more conservative. When a spike in state estimation occurs, the robust controller damps it out quickly rather than aggressively chasing the noisy measurement, preventing erratic flight.
Key Results
- Hardware Success: This is the first successful hardware implementation of a robust, perception-based controller using a data-driven error model.
- Ideal Conditions: In well-lit, feature-rich environments, the $L_1$ robust imitation controller matched the tracking performance of a highly tuned PD controller.
- Degraded Conditions: When tested in simulation with degraded perception (mimicking poor lighting/fewer features), the standard PD controller suffered significantly, whereas the $L_1$ robust controller maintained its tracking and bounded the perception error.
- Bounded Error Guarantees: The perception error stayed well below the theoretical maximum bounds calculated by the SLS framework (1.602 meters in simulation, 3.172 meters in hardware).
Limitations and Future Work
- Overly Conservative Bounds: The theoretical error bounds (e.g., ~3 meters of drift) are technically satisfied but practically unacceptable for precise navigation. The framework currently forces a trade-off that leans too heavily toward minimizing perception error at the expense of tight trajectory tracking.
- Linearization Constraints: The system is linearized around a hover state with a constant yaw. This prevents the drone from executing aggressive maneuvers or actively rotating to face better visual features.
- Future Work: The authors plan to explore Linear Time-Varying (LTV) System Level Synthesis. This would allow them to incorporate yaw into the control model, enabling "active sensing"—where the drone can dynamically turn its camera toward feature-rich areas to actively reduce perception uncertainty.
Summary updated