Evolving Robust, Deliberate Motion Planning With a Shallow Convolutional Neural Network

Abstract

Deep Convolutional Neural Networks (ConvNets) have seen great success on machine learning tasks in recent years but have shown difficulty with tasks that require long-term deliberative planning. Whereas, purpose-built hybrid network architectures have been able to solve increasingly challenging deliberate tasks in two-dimensional and three-dimensional artificial worlds. Starting from a purpose-built network and transitioning to a general architecture, like a deep ConvNet, may retain long-term deliberative planning while allowing greater flexibility in the task domain. This paper employs a standard genetic algorithm (GA) to train the weights of a ConvNet with a recurrent 3x3 filter to produce robust and deliberative motion planning. This technique resulted in an average of 98.97% completion over 10,000 runs in the most difficult deliberate task. This demonstrates that a shallow ConvNet with recurrent connections is capable of producing deliberate and robust motion planning. Further, the evolved ConvNet exhibits superior motion planning in the most challenging environments, when compared to the previous taskspecific motion-planning network.