Abstract
We propose a machine learning enhanced algorithm for solving the optimal landing problem. Using Pontryagin’s minimum principle, we derive a two-point boundary value problem for the landing problem. The proposed algorithm uses deep learning to predict the optimal landing time and a space-marching technique to provide good initial guesses for the boundary value problem solver. The performance of the proposed method is studied using the quadrotor example, a reasonably high dimensional and strongly nonlinear system. Drastic improvement in reliability and efficiency is observed.
Introduction
Landing an aerial vehicle, such as a quadrotor, optimally is a significant challenge in control theory due to the high-dimensional and strongly nonlinear nature of the system dynamics. Traditionally, researchers have relied on two primary approaches: direct methods and indirect methods. Direct methods discretize the problem into a large-scale optimization task, which is robust but often lacks precision. Indirect methods, based on Pontryagin’s Minimum Principle (PMP), derive a Two-Point Boundary Value Problem (TPBVP). While indirect methods yield the most efficient trajectories, they are notoriously sensitive to initial guesses; if the starting trajectory is even slightly inaccurate, the solver fails to converge. This sensitivity has historically limited the practical application of indirect methods in real-world scenarios where initial conditions vary significantly.
Methodology
To overcome the convergence issues of indirect methods, we introduce a hybrid approach that utilizes machine learning to "warm-start" the mathematical solver. The framework consists of three integrated layers:
1. The PMP Framework and Quadrotor Model
The landing problem is transformed into a TPBVP where the objective is to find the optimal state and costate (the mathematical dual of the state) over time. For our study, we utilize a 12-dimensional quadrotor model. The complexity of this system is defined by the various physical parameters and coordinate systems involved.
2. Space-Marching (Homotopy)
Rather than solving a difficult landing from a distant starting point immediately, we employ a space-marching technique. The algorithm solves a sequence of progressively harder problems by picking points along a line from the target back to the starting position. The solution of each "easier" point serves as the initial guess for the next, more distant point, effectively tunneling through the nonlinear landscape.
3. DNN Time Prediction
In "free terminal time" problems, the solver must determine the optimal duration of the flight ($t_f$). This is a critical free parameter; an incorrect guess often leads to solver divergence. We trained a Deep Neural Network (DNN) to predict the optimal landing time based on the initial state. This prediction provides a high-quality anchor for the fixed-time solver.
Experimental Setup and Datasets
Because no standard public benchmark exists for this specific problem, we generated a custom dataset tailored to quadrotor dynamics. Using our space-marching algorithm with a high number of steps to ensure convergence, we collected 300 optimal landing trajectories from random initial states. By sampling 100 points along each path, we created a dataset of 30,000 pairs of "current state" and "remaining optimal time." The final algorithm was evaluated on 100 randomly sampled initial states within a domain of altitudes up to 100 meters.
Key Results
The ML-enhanced algorithm was compared against several baselines using the 12-dimensional quadrotor model:
- Reliability: A standard TPBVP solver with zero initialization achieved a success rate of only 1% to 4%. In contrast, our ML-enhanced algorithm achieved a 99% success rate.
- Efficiency: The average computation time was approximately 17 seconds. While not yet instantaneous, this represents a massive leap in reliability for complex trajectories.
- Optimality: When compared to Model Predictive Control (MPC), which uses a short-horizon approach, our method produced trajectories with significantly lower total cost (fuel and time), as it optimizes over the entire flight path.
Conclusion and Future Work
This research demonstrates that machine learning is most effective when used as a helper for rigorous mathematical solvers rather than a total replacement. By providing accurate initial guesses for $t_f$ and utilizing space-marching, we can unlock the precision of indirect methods for complex landing problems. Future work will focus on incorporating obstacle avoidance into the model and further accelerating the solver to meet the real-time constraints required for on-board deployment.
Summary updated