Approximate Inference-Based Motion Planning by Learning and Exploiting Low-Dimensional Latent Variable Models

Abstract

This work presents an efficient framework to generate a motion plan of a robot with high degrees of freedom (e.g., a humanoid robot). High-dimensionality of the robot configuration space often leads to difficulties in utilizing the widely-used motion planning algorithms, since the volume of the decision space increases exponentially with the number of dimensions. To handle complications arising from the large decision space, and to solve a corresponding motion planning problem efficiently, two key concepts are adopted in this work: First, the Gaussian process latent variable model (GP-LVM) is utilized for low-dimensional representation of the original configuration space. Second, an approximate inference algorithm is used, exploiting through the duality between control and estimation, to explore the decision space and to compute a high-quality motion trajectory of the robot. Utilizing the GP-LVM and the duality between control and estimation, we construct a fully probabilistic generative model with which a high-dimensional motion planning problem is transformed into a tractable inference problem. Finally, we compute the motion trajectory via an approximate inference algorithm based on a variant of the particle filter.

Heterogeneous Dimensionality Reduction for Efficient Motion Planning in High-Dimensional Spaces

“…However, in heterogeneous environments such as the scenarios in Section V, a single-structure dimensionality reduction process may fail or the learned low-dimensional structure may be only locally effective. The results in [5] also showed that the dimensionality of the latent space should be changed based on the planning problem data (different obstacle distributions). In our approach, we propose a DP-based clustering method to find heterogeneous low-dimensional structures for a given planning problem.…”

“…Vernaza and Lee [4] proposed a low-dimensional structure learning method by utilizing the gradient information of the cost function and used a PCA-like method to find the basis of a subspace that contains the start and goal states. Instead of learning the low-dimensional structure from the cost function, Ha et al [5] proposed a method for learning a low-dimensional latent model of a given high-dimensional robot from experts' demonstration data. The robot dynamic model is modeled by a Gaussian process dynamical model.…”

“…The learned structure is directly related to the planning problem rather than simply related to the low-dimensional structure of the robot. In [4], [5], dimensionality reduction were single-structure reduction. However, in heterogeneous environments such as the scenarios in Section V, a single-structure dimensionality reduction process may fail or the learned low-dimensional structure may be only locally effective.…”

“…The low-dimensional structure generally is learned from the cost structure or from experts' demonstration data based on various learning-or modeling-based methods [4], [5]. In those prior works, the dimesionality reduction process learns only a single low-dimensional structure.…”

Fast-Reactive Probabilistic Motion Planning for High-Dimensional Robots

“…Machine learning approaches are still not widely applied in robotic motion planning. Existing applications include guiding the exploration of sampling-based motion planners using nearest neighbor and adaptive sampling [4,5,26], accelerating collision detection through supervised classification [52,53], and pursuing end-to-end motion planning through learning from demonstration [23,61,72]. To the author's best knowledge, this paper is the first application of learning-based methods on the collision risk estimation problem for probabilistic motion planning systems.…”