Tractable robot simulation for terrain leveling

Cook, Daniel (2017) Tractable robot simulation for terrain leveling. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (5MB)

Abstract

This thesis describes the problem of terrain leveling, in which one or more robots or vehicles are used to atten a terrain. The leveling operation is carried out either in preparation for construction, or for terrain reparation. In order to develop and prototype such a system, the use of simulation is advantageous. Such a simulation requires high fidelity to accurately model earth moving robots, which navigate uneven terrain and potentially manipulate the terrain itself. It has been found that existing tools for robot simulation typically do not adequately model deformable and/or uneven terrain. Software which does exist for this purpose, based on a traditional physics engine, is difficult if not impossible to run in real-time while achieving the desired accuracy. A number of possible approaches are proposed for a terrain leveling system using autonomous mobile robots. In order to test these approaches in simulation, a 2D simulator called Alexi has been developed, which uses the predictions of a neural network rather than physics simulation, to predict the motion of a vehicle and changes to a terrain. The neural network is trained using data captured from a high-fidelity non-real-time 3D simulator called Sandbox. Using a trained neural network to drive the 2D simulation provides considerable speed-up over the high-fidelity 3D simulation, allowing behaviour to be simulated in real-time while still capturing the physics of the agents and the environment. Two methods of simulating terrain in Sandbox are explored with results related to performance given for each. Two variants of Alexi are also explored, with results related to neural network training and generalization provided.

Actions (login required)

View Item