Source-linked AI summary
A Survey on Artificial Intelligence Trends in Spacecraft Guidance Dynamics and Control
Dario Izzo, Marcus Märtens, Binfeng Pan
TL;DR
Spacecraft guidance, dynamics, and control increasingly incorporate AI, but evidence spans diverse mission scenarios and techniques. This survey selectively synthesizes mission-motivated work on evolutionary optimisation, tree searches, and machine learning, emphasizing their relationships and synergies. It reports successful applications across trajectory optimisation, landing, and other space-system tasks, while noting that combined techniques can be especially viable for complex challenges.
Problem
AI applications in spacecraft guidance, dynamics, and control are proliferating across diverse scenarios, but individual techniques may perform poorly when transferred between tasks without understanding their strengths and weaknesses.
Method
The survey presents a recent, selective, mission-motivated synthesis of evolutionary optimisation, tree searches, machine learning, deep learning, and reinforcement learning in spacecraft guidance, dynamics, and control.
Results
The surveyed approaches have been successfully applied or are under strong investigation in orbital prediction, planetary landing, spacecraft guidance, interplanetary trajectory optimisation, and low-thrust propulsion.
Takeaways & Limitations
The survey highlights relations and synergies among AI techniques, including combinations that can benefit complex space challenges.
Abstract
from arXiv · showhide
The rapid developments of Artificial Intelligence in the last decade are influencing Aerospace Engineering to a great extent and research in this context is proliferating. We share our observations on the recent developments in the area of Spacecraft Guidance Dynamics and Control, giving selected examples on success stories that have been motivated by mission designs. Our focus is on evolutionary optimisation, tree searches and machine learning, including deep learning and reinforcement learning as the key technologies and drivers for current and future research in the field. From a high-level perspective, we survey various scenarios for which these approaches have been successfully applied or are under strong scientific investigation. Whenever possible, we highlight the relations and synergies that can be obtained by combining different techniques and projects towards future domains for which newly emerging artificial intelligence techniques are expected to become game changers.
1 Introduction
This survey reviews recent AI developments in spacecraft guidance, dynamics, and control, emphasizing mission-relevant applications and combinations of evolutionary optimisation, tree searches, and machine learning. It deliberately presents a selective, recent, practically motivated overview rather than a comprehensive historical account.
- AI applications in the space sector span spacecraft design, mission operations, guidance and control, navigation, perturbed-motion prediction, and astronomical-object classification.
- The survey focuses on recent work in spacecraft guidance, dynamics, and control that is motivated by established or proposed mission concepts.The authors exclude strongly speculative publications and prioritize work from the last few years.
- The survey organizes its discussion around evolutionary optimisation, tree searches, and machine learning, including deep learning and reinforcement learning.It also highlights relations and synergies among these techniques.
- Results from one AI technology may transfer poorly across tasks when its strengths and weaknesses are not understood, motivating attention to benchmarks and combined techniques.The survey cites pioneering work and competitions such as GTOC as sources of improved understanding.
- The report presents a timely state-of-the-art view of AI for guidance and control, while related surveys address distributed spacecraft AI, geoscience, remote sensing, and historical perspectives.
2 Evolutionary Optimisation
Evolutionary optimisation methods address continuous, multi-objective, and combinatorial spacecraft-design problems. The survey describes algorithm-specific applications, performance dependencies, hybrid strategies, and automated trajectory pipelines.
- Single-objective, Unconstrained, Continuous Problems: Evolutionary algorithms use heuristic population-based search and are useful for interplanetary trajectory optimisation in discontinuous and rugged solution landscapes.Standard genetic algorithms apply mutation, crossover, and selection to populations of candidate solutions.
- Single-objective, Unconstrained, Continuous Problems: Differential Evolution is effective for nonlinear, non-differentiable continuous spacecraft-transfer problems, while self-adaptive variants evolve algorithm hyper-parameters with solutions.
- Single-objective, Unconstrained, Continuous Problems: PSO, DE, and related meta-heuristics have problem-dependent performance, encouraging parallel evaluation or combinations such as DE–PSO for automated interplanetary mission design.
- Single-objective, Unconstrained, Continuous Problems: CMA-ES adapts mutation using decision-variable covariance and has been shown to outperform DE or PSO on a large class of interplanetary transfer problems.
- Multi-objective Problems: Multi-objective and combinatorial mission problems require Pareto-front methods, hybrid search, or automated decomposition of encounter sequences and inner low-thrust optimisation.Examples include NSGA-II, MOEA/D, genetic algorithms, monotonic basin hopping, and travelling-salesman formulations.
3 Tree Searches
Tree searches address large combinatorial spacecraft-design spaces by incrementally expanding solution nodes while selectively exploring promising branches. Beam Search, MCTS, and hybrid variants provide different exploration, stochasticity, and resource-quality trade-offs.
- Tree searches are appropriate when solutions can be constructed incrementally from smaller separable sub-problems, avoiding ineffective sampling of very large combinatorial spaces.
- Tree-search nodes represent decision points such as the next orbital body, while search strategies limit expansion because exhaustive enumeration is intractable.
- Beam Search performed best among studied strategies, especially for massive search spaces, and balances exploration with exploitation by ranking and retaining a subset of nodes.
- Lazy-Race Tree Search ranks partial trajectories by total flight time while expanding nodes to minimize mass consumption, unlike a greedy level-wise strategy.
- MCTS rediscovered the Cassini planetary-encounter sequence within a timeline close to the flown trajectory, while Beam P-ACO offers anytime solutions through pheromone-guided exploration.
4 Machine Learning
The survey reviews machine-learning applications in spacecraft guidance, dynamics, and control, emphasizing synergies with evolutionary optimisation and emerging deep and reinforcement learning methods. Reported applications include trajectory optimisation, orbit prediction, visual control, rendezvous, and planetary landing.
- Machine learning adoption in spacecraft guidance and control remains limited by scarce suitable large-scale aerospace datasets and uncertain applicability.
- 4.1 Machine Learning and Evolutionary Optimisation: Evolutionary optimisation can generate training data, provide initial guesses, construct surrogate objectives, or combine with learning models to accelerate search.Optimisation produces assessed design points that can train supervised models; learned initial conditions can reduce later evolutionary search effort.
- 4.2 Deep Learning: Deep networks learn optimal spacecraft manoeuvres, including Earth-to-Mars transfers, and may generalize beyond training bounds while supporting analysis of stability margins.Guidance and Control Networks use imitation learning and differential algebra with automated differentiation to study stability and control performance.
- 4.2 Deep Learning: Deep-learning applications also use visual cues, Gaussian-process regression, random forests, and SVMs for control, asteroid transfers, manifold approximation, and orbit classification or prediction.A Gaussian-process model reached high accuracy from 300 numerically derived samples after grouping asteroids by similar orbits; random forests performed reasonably across multiple orbital energies.
- 4.4 Reinforcement Learning: Reinforcement-learning studies address rendezvous, collision avoidance, planetary landing, and robust low-thrust control in uncertain environments.DQN controllers learn obstacle avoidance and target seeking, while policy-gradient methods support fuel-optimal pinpoint landing under varying constraints.
- 4.4 Reinforcement Learning: A visual-feedback controller improved accuracy by an order of magnitude, transferred to a more general gravitational model, and used neuroevolutionary PSO training.
5 Final Remarks
The survey reports that AI methods have produced competitive and often superior approaches across several spacecraft applications. It expects further developments in autonomous space-system tasks, alongside validation methods to increase trust in trained models.
- 5 Final Remarks: AI methods have been fruitful in orbital prediction, planetary landing, spacecraft guidance, interplanetary trajectory optimisation, and low-thrust propulsion.The reported methods and architectures are described as competitive, and often superior, to current state-of-the-art methodologies.
- 5 Final Remarks: Future success stories are expected in formation flying, rendezvous and docking, in-orbit self-assembly, and autonomous surface detection.The survey specifically anticipates Deep Learning and Deep Reinforcement Learning powering many such developments.
- 5 Final Remarks: Validation work is expected to develop methods that increase trust in trained models such as deep neural networks.