自由漂浮空间机器人轨迹采样及其多模态 优化方法

Motion planning，critical for space robots，is formulated as a trajectory optimization problem. Conventional methods often neglect multimodality in non-convex cost functions，limiting diverse motion strategy generation. A multimodal optimization approach based on sampled trajectory clustering is proposed，enabling multiple feasible solutions with similar costs in a single planning iteration. Joint trajectories are parameterized using Bézier curves and modeled as non-convex optimization problems. Cost function multimodal search is transformed into trajectory clustering via importance weights. A variational autoencoder（VAE）approximates optimal trajectory distributions under collision-free conditions， enhancing initialization efficiency and quality. Gaussian mixture model parameters are estimated through variational Bayesian expectation-maximization，facilitating trajectory clustering and local optimization. Simulation results confirm the method effectively captures cost function multimodality，providing redundant solutions for space robot motion planning and overcoming limitations of conventional approaches.