A Constrained Motion Planning Method Exploiting Learned Latent Space for High-Dimensional State and Constraint Spaces*

This article presents a novel approach to address high-dimensional constrained motion planning problems. The proposed method exploits learned latent spaces to efficiently find a feasible constrained path. Although recent data-driven methods have reduced planning time, two major challenges arise as the space dimensionality increases, such as in multi-arm manipulation. First, the configuration space search employed by existing data-driven approaches becomes computationally expensive as the complexity of constraints increases. Second, preparing datasets for high-dimensional problems is a time-consuming task. To address these challenges, this article introduces a novel approach: the latent motion method. Instead of exploring the configuration space, the latent motion method explores the latent space with a latent jump method that mitigates topological problems. In addition, a tangent space dataset augmentation technique is employed to approximate the manifold using a sparse dataset. Experimental evaluations on benchmark problems indicate that the proposed approach outperforms existing methods in high-dimensional constrained motion planning problems.