TL;DR
This paper introduces NeuroMHE, a novel neural network-enhanced Moving Horizon Estimator for quadrotors. NeuroMHE automatically tunes its internal parameters using a neural network, allowing it to adapt to various flight conditions and disturbances without requiring extensive ground-truth data. This approach significantly improves disturbance estimation accuracy and enables more robust flight control.
What Problem Does This Paper Address?
Quadrotors are increasingly used for complex tasks, but they are susceptible to various disturbances like wind, unmodeled aerodynamics, and external forces. Effectively estimating and reacting to these disturbances is crucial for maintaining stable and accurate flight. Traditional estimation methods often require extensive manual tuning for specific scenarios or large datasets of ground-truth disturbance information, which are difficult to obtain. This paper addresses the challenge of creating a disturbance estimator that is both highly accurate and adaptable to diverse and changing flight conditions, while minimizing the need for specialized training data.
Methods
The core innovation is the NeuroMHE (Neural Moving Horizon Estimator). It integrates a Moving Horizon Estimator (MHE) with a portable neural network.
- MHE: This is a standard technique for state estimation that solves a nonlinear optimization problem over a sliding window of past measurements and control inputs. Its performance heavily relies on carefully chosen weighting matrices in its cost function.
- Neural Network for Adaptive Weightings: Instead of manually tuning the MHE's weighting matrices, NeuroMHE uses a neural network to generate these parameters online. This allows the estimator to adapt its behavior based on the current flight state or other relevant signals.
- Analytical Gradients via Kalman Filter: A key technical contribution is the derivation of analytical gradients of the MHE estimates with respect to the neural network's parameters. Crucially, these gradients are computed efficiently using a recursive Kalman filter-based solver. This avoids computationally expensive matrix inversions and enables seamless integration of the MHE as a learnable layer within the neural network.
- Model-Based Policy Gradient Training: NeuroMHE is trained using a model-based policy gradient algorithm. This allows the neural network parameters to be learned directly from the quadrotor's trajectory tracking error, eliminating the need for ground-truth disturbance data.
- Sparse Parameterization: To keep the neural network portable and efficient, the paper proposes a sparse parameterization of the weighting matrices, reducing the number of learnable parameters.
- Constrained MHE: The framework is extended to handle inequality constraints, improving safety by preventing unrealistic estimations.
Datasets
- Agile and Extreme Flight Dataset: Used for evaluating NeuroMHE against the state-of-the-art estimator NeuroBEM. This dataset was collected in a motion capture room and includes ground-truth disturbance data.
- Synthetic Disturbance Datasets: Used for training and evaluating NeuroMHE's ability to adapt to dynamic noise covariances and for testing its performance on unseen trajectories.
- Real-world Flight Data: Collected from physical experiments on quadrotors under various challenging conditions (cable forces, downwash flow).
Key Insights and Ideas
- Adaptive Weightings are Key: The paper demonstrates that dynamically adjusting MHE's weighting matrices based on the flight context is crucial for robust performance, especially when noise characteristics change.
- Kalman Filter for Gradient Efficiency: Using a Kalman filter to compute gradients for MHE optimization is a clever way to achieve computational efficiency, making the approach practical for real-time applications.
- Learning from Tracking Error is Sufficient: The ability to train the disturbance estimator directly from trajectory tracking error, rather than requiring ground-truth disturbance data, significantly lowers the barrier to entry and makes the method more broadly applicable.
- Portability and Efficiency: The proposed neural network architecture and sparse parameterization result in a surprisingly small and portable model that still achieves superior performance.
Key Results
- Improved Force Estimation: NeuroMHE reduces force estimation errors by up to 76.7% compared to the state-of-the-art NeuroBEM.
- Parameter Efficiency: NeuroMHE uses a neural network with only 7.7% of the learnable parameters of NeuroBEM.
- Efficient Training: NeuroMHE can be trained effectively in a few episodes, even from scratch, using trajectory tracking error.
- Robust Control Performance: In simulations, NeuroMHE reduces average tracking errors by up to 86.3% compared to fixed-weighting MHE and L1 adaptive control.
- Physical Experiment Validation: NeuroMHE successfully robustifies a baseline flight controller against challenging disturbances like state-dependent cable forces and downwash flow, outperforming other methods in terms of estimation and tracking accuracy.
Limitations and Future Work
- Computational Resources: While efficient, NeuroMHE still requires solving a nonlinear optimization problem, demanding more computational resources than simpler methods. Running at 25 Hz for position control is feasible, but higher frequencies for torque control might require more powerful hardware or optimized solvers.
- Stability Guarantees: The paper acknowledges that theoretical stability guarantees for nonlinear MHE and neural network-based estimators are still an open research problem, unlike methods like L1-AC.
- Future Work: The authors suggest developing more efficient training algorithms and further analyzing the stability of NeuroMHE. They also mention the potential for applying the method to other robotic systems.
Summary updated