Robot Motion Planning in Learned Latent Spaces

Abstract

This paper presents Latent Sampling-based Motion Planning (L-SBMP), a methodology towards computing motion plans for complex robotic systems by learning a plannable latent representation. Recent works in control of robotic systems have effectively leveraged local, low-dimensional embeddings of high-dimensional dynamics. In this paper we combine these recent advances with techniques from samplingbased motion planning (SBMP) in order to design a methodology capable of planning for high-dimensional robotic systems beyond the reach of traditional approaches (e.g., humanoids, or even systems where planning occurs in the visual space). Specifically, the learned latent space is constructed through an autoencoding network, a dynamics network, and a collision checking network, which mirror the three main algorithmic primitives of SBMP, namely state sampling, local steering, and collision checking. Notably, these networks can be trained through only raw data of the system's states and actions along with a supervising collision checker. Building upon these networks, an RRT-based algorithm is used to plan motions directly in the latent space -we refer to this exploration algorithm as Learned Latent RRT (L2RRT). This algorithm globally explores the latent space and is capable of generalizing to new environments. The overall methodology is demonstrated on two planning problems, namely a visual planning problem, whereby planning happens in the visual (pixel) space, and a humanoid robot planning problem.

LEGO: Leveraging Experience in Roadmap Generation for Sampling-Based Planning

“…We consider samplers that do not assume offline computation or learning such as Medial-Axis PRM (MAPRM) [6,14], Randomized Bridge Sampler (RBB) [15], Workspace Importance Sampler (WIS) [16], a Gaussian sampler, GAUS-SIAN [20], and a uniform Halton sequence sampler, HALTON [4]. Additionally, we also compare our framework against the state-of-the-art learned sampler SHORTESTPATH [35] upon which our work is based.…”

“…There has been a lot of recent effort on finding low dimensional structure in planning [31]. In particular, generative modeling tools like variational autoencoders [32] have been used to great success [33][34][35][36][37]. We base our work on Ichter et al [11] where a CVAE is trained to learn the shortest path distribution.…”

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

“…The robot dynamic model is modeled by a Gaussian process dynamical model. The authors in [22] also proposed using an autoencoding network to learn the low-dimensional latent dynamic model and then searching an optimal trajectory in the latent space, ultimately decoding the low-dimensional trajectory to one in the full state space using the trained decoder. Although the lower dimensional dynamical representation learning is a promising solution for achieving the optimal control of complex high-dimensional robots, it is inefficient for planning because the training process is lengthy and requires large amounts of training data.…”

Harnessing Reinforcement Learning for Neural Motion Planning

“…The idea of learning from experience to plan faster in motion planning has been studied under various approaches, such as using a library of past trajectories [7,20,6,10], learning the sampling procedure [32,19,25], learning a latent representation of obstacles [19], and learning to select goals [12].…”