Neural Network Approach to Feature Sensitive Motion Planning
Abstract
Motion planning (MP) is the problem of finding a valid path (e.g., collision free) from a start to a goal state for a movable object. MP is a complex problem with a myriad of applications, ranging from robotics, to computer-aided design, to computational biology. Sampling-based planning deals with MP’s complexity by constructing a graph which approximates the planning space. Different sampling based planners have been developed to tackle specific scenarios, but none of these is best for every scenario, e.g., cluttered vs. free space vs narrow passage. Thus, adaptive methods were created to combine different samplers effectively to solve more complex and heterogeneous environments.
Adaptive methods have been proposed that learn the best sampler for the entire space or that partition the space into simple and discrete region types, which are suited for particular samplers. These methods do not solve the problem of environments containing multiple complex areas that are difficult to automatically partition. In this thesis, we propose an alternative approach using neural networks to create an adaptive method that does not require regions. We replace the concept of regions with a visibility distribution, how “free” a node is, allowing our method to work for a wider range of interesting problems. Experiments show significant improvement in speed compared to methods that attempt to use a single sampler for a complex environment.
Subject
motionplanning
neural
network
networks
algorithms
AI
artificial
intelligence
parasol
robotics
Citation
Medina Vargas, Jose Andres (2014). Neural Network Approach to Feature Sensitive Motion Planning. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /157594.