Low-dimensional projections for SyCLoP

Abstract

This paper presents an extension to SyCLoP, a multilayered motion planning framework that has been shown to successfully solve high-dimensional problems with differential constraints. SyCLoP combines traditional sampling-based planning with a high-level decomposition of the workspace through which it attempts to guide a low-level tree of motions. We investigate a generalization of SyCLoP in which the highlevel decomposition is defined over a given low-dimensional projected subspace of the state space. We begin with a manually-chosen projection to demonstrate that projections other than the workspace can potentially work well. We then evaluate SyCLoP's performance with random projections and projections determined from linear dimensionality reduction over elements of the state space, for which the results are mixed. As we will see, finding a useful projection is a difficult problem, and we conclude this paper by discussing the merits and drawbacks of various types of projections.

Region-Guided and Sampling-Based Tree Search for Motion Planning With Dynamics

“…In particular, the experiments use a nonlinear ground vehicle model, a high-dimensional nonlinear second-order snake-like robot model, and a nonlinear second-order aerial-vehicle model operating in complex environments. Statistical analysis is conducted to measure the performance of GUST and compare it to related work, e.g., Syclop [32], VSyclop 1 [46], KPIECE [29], PDST [28], RRT [23], fRRT [30], TRRT [27], and STRIDE [31]. Statistical analysis is also conducted to measure the impact of the decomposition.…”

“…GUST is compared with Syclop [32], VSyclop [46], KPIECE [29], PDST [28], RRT [23], fRRT [30], TRRT [27], STRIDE [31], and also with the preliminary version of GUST [43]. The implementation of Syclop and VSyclop includes several features such as switching between shortest-path and random guides and abandoning the current guide when little progress is made.…”