Fastron: A Learning-Based Configuration Space Model for Rapid Collision Detection for Gross Motion Planning in Changing Environments

Proc. Robotics: Science and Systems Workshop on (Empirically) Data-Driven Manipulation, 2017

Nikhil Das, Naman Gupta, Michael Yip

Abstract: A. Contributions Robot manipulation requires both gross and fine motion planning. Realizing the high cost involved in kinematic-based collision detections (KCDs) for motion planning, we present a fast technique to update an approximate C-space representation to be used as a proxy collision detector for gross motion planning in this paper. The purpose of this effort is to reduce the computation required for gross motion planning such that more resources may be dedicated toward sampling or data collection required for fine motion planning. The novel contributions of this work are: 1) a simple yet efficient method to sparsely represent C-space obstacles using a kernel perceptron hyperplane 2) a modified kernel perceptron that allows addition and removal of support points, and 3) an active learning strategy to update the model in response to a changing environment.

Das et al. (2017) Fastron: A Learning-Based Configuration Space Model for Rapid Collision Detection for Gross Motion Planning in Changing Environments, Proc. Robotics: Science and Systems Workshop on (Empirically) Data-Driven Manipulation.

Pub Link: http://ddm2017.mit.edu/sites/default/files/documents/10.pdf
arXiv: