Louis Ly: Autonomous Exploration, Reconstruction, and Surveillance of Topologically-Complex Environments using Deep Learning
Time: Thu 2019-02-14 14.15 - 15.00
Lecturer: Louis Ly, University of Texas at Austin
Location: Room F11, Lindstedtsvägen 22, våningsplan 2, F-huset, KTH Campus.
Consider the problem of generating a minimal sequence of observing locations to achieve complete line-of-sight visibility coverage of an environment. If the environment is initially unknown, the problem is called exploration and reconstruction. This is particularly useful for autonomous agents to map out unknown, or otherwise unreachable environments, such as undersea caverns. If the environment is known, the problem is one of surveillance: how should a minimal set of sensors be placed to maintain complete surveillance of an environment? In this talk, I will introduce a level-set approach which greedily determines vantage points by optimizing a notion of visibility information gain. By leveraging recent advances in machine learning, namely convolutional neural networks, we can efficiently approximate the amount of information gain at each potential vantage point, even when the environment is unknown. I'll present simulations in complex 3D urban environments and also discuss a future extension to multi-step path planning.