TL;DR
This paper successfully implements a data-enabled predictive control (DeePC) algorithm for real-world quadcopter trajectory tracking. By leveraging past input-output data, DeePC bypasses traditional system identification, demonstrating robust performance even with nonlinear dynamics and noisy measurements, and proving computationally tractable for real-time control.
What Problem Does This Paper Address?
Traditional model-based control approaches for systems like quadcopters rely heavily on accurate system identification, a process that is often time-consuming, complex, and requires expert knowledge. This can be particularly challenging for nonlinear systems with noisy measurements. While data-driven methods offer an alternative, many still require learning an explicit input-output model, which can introduce its own set of tuning challenges. This paper addresses the need for a control method that can directly use system data to predict future behavior and optimize control actions without explicit model identification or learning, specifically for the challenging case of a real-world quadcopter.
Methods
The core of this work is the application and validation of the Data-Enabled Predictive Control (DeePC) algorithm. DeePC is a finite-horizon optimal control method that uses historical input-output data to form "motion primitives." These primitives are then linearly combined to predict future system trajectories. A key contribution is the use of a regularized variant of DeePC to handle the nonlinear dynamics and noisy measurements inherent in real-world quadcopters. This regularization is incorporated directly into the optimization objective. The paper also details a data collection procedure involving injecting a pseudorandom binary sequence (PRBS) signal to ensure the data is persistently exciting, a requirement for the algorithm. For comparison, a model-based Model Predictive Control (MPC) approach using a linearized, first-principles model of the quadcopter is also implemented.
Datasets
The paper utilizes data collected from real-world experiments with a Crazyflie 2.0 quadcopter. This data consists of input (thrust, body rate references) and output (position, orientation) measurements. For simulation purposes, a nonlinear model of the quadcopter is used, with measurements augmented by simulated zero-mean Gaussian noise. No new benchmark datasets are introduced; rather, the paper focuses on applying DeePC to a real-world system.
Key Insights and Ideas
- Data as a Model: DeePC fundamentally treats historical input-output data as a direct representation of the system's dynamics, bypassing the need for explicit model building.
- Regularization is Crucial for Nonlinearity and Noise: For real-world systems like quadcopters, which are nonlinear and subject to measurement noise, regularization is essential. Without it, the algorithm can produce physically unreasonable predictions, leading to poor performance.
- Hyperparameter Tuning is Key: The performance of DeePC is sensitive to hyperparameters such as the amount of data collected ($T_a$), the initial condition estimation horizon ($T_{ini}$), and regularization weights ($\lambda_s$, $\lambda_g$). Simulation-based analysis is vital for understanding these sensitivities and deriving tuning guidelines.
- Transferability from Simulation to Reality: The insights gained from simulation-based hyperparameter tuning largely transferred to the real-world quadcopter experiments, requiring minimal adjustments. This suggests the robustness of the DeePC approach.
- Computational Tractability: The DeePC algorithm, even with its optimization problem, is shown to be computationally feasible for real-time control on the quadcopter, with solve times well within the required limits.
Key Results
- The DeePC algorithm achieved successful trajectory tracking for the quadcopter, including step changes and figure-8 patterns.
- The regularized DeePC demonstrated robustness to nonlinear dynamics and noisy measurements, outperforming an unregularized version.
- Simulation-based analysis provided valuable insights into hyperparameter selection, leading to effective tuning guidelines.
- Experimental results validated the simulation findings, showing similar performance and trajectory behavior.
- Comparison with a model-based MPC showed qualitatively similar performance for vertical position tracking, with DeePC exhibiting a faster rise time in the horizontal plane and a better ability to compensate for steady-state input mismatches.
- The algorithm was computationally tractable, with solve times for the optimization problem being significantly faster than the real-time requirement (e.g., ~4-6 ms for $p=3$ and $p=5$ respectively).
Limitations and Future Work
- Model Mismatch Robustness: While DeePC shows some robustness to model mismatch (e.g., yaw angle calibration error), further investigation is needed to understand the extent of this capability.
- Hyperparameter Generalization: While $\lambda_g$ and $q$ are important, the paper notes that general guidelines for their selection across different systems are still an area for future research, though $\lambda_g = 500$ and $q=2$ appear to be a good starting point.
- Online Data Updates: The paper suggests exploring the concept of updating the Hankel matrix online to adapt to slowly drifting measurement calibrations, rather than requiring periodic recalibration as in model-based approaches.
- Application to Other Systems: The authors propose applying DeePC to other real-world systems for which first-principles models are difficult or impossible to derive.
Summary updated