As space missions become increasingly complex, tasks such as on-orbit servicing, deepspace exploration, and extraterrestrial construction require highly autonomous and reliable robotic systems. Artificial intelligence (AI) has emerged as a key enabling technology for perception, decision-making, and autonomous control in space robotics. However, space environments impose severe constraints, including limited computation, sparse data, radiation, and long communication delays, motivating the development of sustainable AI systems that are efficient, robust, and adaptive over long-duration missions. This paper reviews recent advances in sustainable AI for space robotics, focusing on deep learning, reinforcement learning (RL), physics-informed neural networks (PINNs), and large language models (LLMs). Deep learning methods, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GNNs), have shown strong performance in perception, trajectory prediction, and multi-agent coordination. RL enables autonomous control and decision-making in dynamic environments, while PINNs improve physical consistency and data efficiency by incorporating governing laws into the learning process. LLMs further enhance high-level reasoning, mission planning, and human–robot interaction capabilities. Finally, we discuss current challenges, including data scarcity, sim-to-real transfer, computational limitations, and long-term autonomy, and outline future directions for sustainable AI in next-generation space missions.
Accurate rock segmentation on the Martian surface is essential for autonomous path planning and safe rover navigation. This paper presents WaveMars-LiteNet, a lightweight semantic segmentation network specifically designed to address the challenges of low grayscale contrast, substantial target scale variation, and stringent computational constraints inherent in onboard planetary exploration missions. The proposed architecture integrates a multi-branch backbone with three core components: a dual-channel feature extraction module (DFEM) that enhances edge and texture discrimination, a wavelet dual-channel downsampling module (WDCDM) that fuses spatial and frequency-domain information for efficient feature reduction, and a wavelet multi-texture fusion attention module (WTFAM) that strengthens fine-grained detail representation through frequency-aware attention mechanisms. To evaluate performance under realistic operational conditions, we introduce the ZRMars-Rocks dataset, comprising annotated imagery acquired by the Zhurong Mars rover with multi-scale augmentation. Comprehensive experiments on real-world datasets (ZRMars-Rocks and MarsData-V2) and simulated benchmarks (SimMars6K and SynMars-8k) demonstrate competitive performance against representative lightweight and high-complexity segmentation models. With merely 2.38M parameters and 7.91 giga floating-point operations (GFLOPs), WaveMars-LiteNet achieves 81.68
Future exploration of the Moon will involve numerous both crewed and uncrewed launches, essential for transporting materials and instruments critical to the success of lunar missions. The L1 Lagrangian point can represent an ideal transit hub for space missions due to its dynamical properties and strategic location between Earth and Moon. This paper shows an efficient and systematic technique based on the theory of functional connections to evaluate transfers from Earth to Moon, making use of the stable/unstable manifolds associated with the family of Lyapunov orbits around L1. Tens of millions of transfers are evaluated, narrowing the space of solutions to the ones with the lowest costs in an extensive investigation. The transfer is designed under the circular restricted three-body problem formulation with a specific system of coordinates and constraints. A Moon flyby is followed by an entry into a Lyapunov orbit around L1, from where a final transfer to the Moon is performed, with every leg aided by an invariant manifold. The results show a reduced fuel consumption of at least 58.80 m/s as compared to the literature. This difference can be useful for mission designers, while the efficiency of the technique can potentially extend the results to other types of transfers in space.
The increase in human space activities has increased space threats, such as debris and failed spacecraft, which has made space orbital game technology a critical tool for ensuring space security. Although differential game theory remains the mainstream approach for classic space orbital games, its application faces growing limitations because of nonlinear orbital dynamics, model complexity, computational intractability, and poor real-time performance in increasingly complex scenarios. To address these challenges, numerous extended orbital game models have been developed in recent years, along with new solution methods. This review provides a systematic analysis and summary of model extensions and solution methods in orbital games. First, the extension directions of space orbital games are analyzed, and five extended game models based on the basic orbital game model are summarized: the closed-loop guidance model, incomplete information model, diverse scenario game model, impulsive thrust model, and multispacecraft game model. Afterward, research achievements in solution methods, including qualitative differential games, quantitative differential games, and artificial intelligence methods, are categorized and reviewed. Finally, current challenges and potential future research directions are discussed.
This paper presents an improved two-phase sequential convex programming (SCP) algorithm for reentry trajectory optimization. The continuous-time optimal control problem is first discretized using the Chebyshev pseudospectral method (CPM). A conformal mapping is introduced to redistribute the Chebyshev–Gauss–Lobatto (CGL) nodes, which effectively mitigates the ill-conditioning of the differentiation matrix and enhances numerical stability. The influence of the trust-region radius on the descent behavior of the penalty function is then analyzed, leading to a two-phase optimization strategy. In phase I, line-search is adopted to rapidly obtain a feasible solution, thereby avoiding convergence delays caused by prematurely large penalty coefficients. In phase II, the step size is finely tuned using trust-region constraints, ensuring both convergence accuracy and improved robustness. Numerical simulations demonstrate that the proposed algorithm achieves superior performance in terms of convergence speed and solution accuracy, while effectively eliminating the numerical oscillations typically induced by excessive penalty coefficients in conventional penalty methods. The framework provides a reliable and efficient numerical approach for reentry trajectory optimization.
