Machine learning and evolutionary techniques in interplanetary trajectory design

Dario Izzo, Christopher Iliffe Sprague, Dharmesh TailorarXiv 2018 · OpenReview

Abstract

After providing a brief historical overview on the synergies between artificial intelligence research, in the areas of evolutionary computations and machine learning, and the optimal design of interplanetary trajectories, we propose and study the use of deep artificial neural networks to represent, on-board, the optimal guidance profile of an interplanetary mission. The results, limited to the chosen test case of an Earth-Mars orbital transfer, extend the findings made previously for landing scenarios and quadcopter dynamics, opening a new research area in interplanetary trajectory planning.

Introduction

The design of interplanetary trajectories has historically relied on a synergy between artificial intelligence and optimal control theory. For decades, evolutionary algorithms—such as Genetic Algorithms and Particle Swarm Optimization—have served as the "old guard," effectively pruning the massive search spaces required to find viable paths through the solar system. However, as mission requirements move toward greater autonomy, a new frontier has emerged: the use of Deep Learning to represent optimal guidance laws directly on-board the spacecraft.

The Challenge of On-Board Optimization

Designing a trajectory from Earth to Mars is a monumental mathematical task. While ground-based supercomputers can solve these complex "optimal control" problems to find fuel-efficient paths, executing these calculations in real-time on a spacecraft is significantly more difficult.

Space-qualified processors are designed for reliability and radiation hardening rather than raw speed; they are often orders of magnitude slower than standard consumer hardware. Consequently, a single optimization iteration can take nearly a second—far too slow for the dynamic adjustments required during flight. This research proposes using Deep Neural Networks (DNNs) as "surrogate models." By training a network to learn the mapping between a spacecraft's state and its optimal thrust, the system can provide guidance commands via a single, computationally efficient forward pass.

Methodology

The study employs a three-step pipeline to transition from mathematical theory to a deployable neural model:

  1. Data Generation: Using Pontryagin’s Maximum Principle (PMP), the researchers solved thousands of optimal control problems to create a "ground truth" dataset. This dataset maps spacecraft states (position, velocity, and mass) to their corresponding optimal thrust vectors.
  2. Supervised Learning: A deep feed-forward neural network was trained to perform regression on this data, effectively learning to "remember" the optimal solution for any given state within the training distribution.
  3. Architecture: The final model consists of four hidden layers with 200 units per layer. It utilizes ReLU activation functions for hidden layers and a Tanh activation for the output layer to ensure thrust commands remain within physical limits.

Datasets and Control Laws

The research focuses on an Earth-Mars low-thrust transfer, utilizing two distinct datasets to test the network's capabilities:

  • $T_{QOC}$ (Quadratic Optimal Control): This dataset represents trajectories with "smoothed" quadratic control laws. These result in continuous thrust profiles that are mathematically easier for a neural network to approximate.
  • $T_{MOC}$ (Mass Optimal Control): This represents the fuel-minimization goal essential for real-world missions. These trajectories utilize "bang-bang" control, where engines switch abruptly between 0% and 100% thrust. This dataset contains over 308,000 state-control pairs but presents a significant challenge due to these sharp discontinuities.

Key Results and Insights

The study demonstrates that treating optimal guidance as a regression problem is a viable path toward autonomous flight.

  • Accuracy in Smooth Scenarios: For the $T_{QOC}$ trajectories, the network achieved high precision, with a Mean Squared Error (MSE) as low as 0.0075 for throttle magnitude. Qualitative analysis showed that predicted thrust angles almost perfectly overlapped with the ground truth.
  • The Discontinuity Barrier: The primary difficulty remains the "bang-bang" control in mass-optimal scenarios. The DNN tends to "blur" the switching points where the engine should instantly engage or disengage, leading to higher error rates at those specific temporal moments.
  • Computational Efficiency: The core advantage of this approach is its compatibility with radiation-hardened hardware. A DNN forward pass requires fixed, predictable CPU cycles, making it a robust alternative to iterative solvers that may fail to converge within strict time windows.

Limitations and Future Work

While the results are promising, several hurdles remain before deep learning can take the helm of an interplanetary mission. The current models struggle with the high-frequency switching of mass-optimal engines; future research may require specialized architectures, such as Transformers or LSTMs, to better handle these temporal discontinuities.

Furthermore, this study focused on "open-loop" accuracy—measuring how well the network predicts a single point in time. The next critical step is "closed-loop" validation, where the neural network is responsible for flying a simulated spacecraft from start to finish. This will determine if small prediction errors accumulate into significant trajectory deviations over the course of a multi-month mission. Finally, while this study focused on Earth-Mars transfers, the long-term goal remains the development of generalized models capable of navigating between any planetary bodies.

Summary updated

Abstract

After providing a brief historical overview on the synergies between artificial intelligence research, in the areas of evolutionary computations and machine learning, and the optimal design of interplanetary trajectories, we propose and study the use of deep artificial neural networks to represent, on-board, the optimal guidance profile of an interplanetary mission. The results, limited to the chosen test case of an Earth-Mars orbital transfer, extend the findings made previously for landing scenarios and quadcopter dynamics, opening a new research area in interplanetary trajectory planning.
BibTeX
@misc{izzo2018machine,
  author = {Dario Izzo and Christopher Iliffe Sprague and Dharmesh Tailor},
  title = {{Machine learning and evolutionary techniques in interplanetary trajectory design}},
  year = {2018},
  eprint = {1802.00180},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.00180},
}
Figure