Accelerating for Trajectory Generation in Spacecraft Proximity via MLP-LSTM Network

Trajectory generation is fundamental and crucial technology in spacecraft missions. As the safety-crucial constraints are generally complex, AI-based methods have demonstrated superior performance compared to traditional numerical approaches. However, as the number of spacecraft missions increases, the computation resources allocated to trajectory generation are limited, and most current AI-based methods require a large amount of computational resources. Therefore, achieving trajectory generation with limited computational resources remains a significant research challenge. This paper proposes an MLP-LSTM network architecture by modifying the basic LSTM network to address the mentioned limitation. This architecture extracts informative features from the observed data in the past time period through a multilayer perceptron, and uses the LSTM network to predict the control output and form an open-loop trajectory. Then, a sequence solver for a non-convex optimization problem is warm-started to obtain the trajectory satisfying the safety constraints. Simulation results demonstrate that, compared to state-of-the-art Transformer-based architectures, the proposed architecture can effectively reduce the memory size of network model parameters, improve solving efficiency, and even support data inference on the CPU.