Learning navigation policies in an unstructured terrain is a complex task. The Learning to Search (LEARCH) algorithm constructs cost functions that map environmental features to a certain cost for traversing a patch of terrain. These features are abstractions of the environment, in which trees, vegetation, slopes, water and rocks can be found, and the traversal costs are scalar values that represent the difficulty for a robot to cross given the patches of terrain. However, LEARCH tends to forget knowledge after new policies are learned. The study demonstrates that reinforcement learning and long-short-term memory (LSTM) neural networks can be used to provide a memory for LEARCH. Further, they allow the navigation agent to recognize hidden states of the state space it navigates. This new approach allows the knowledge learned in the previous training to be used to navigate new environments and, also, for retraining. Herein, navigation episodes are designed to confirm the memory, learning policy and hidden-state recognition capabilities, acquired by the navigation agent through the use of LSTM.
Part of the book: Advanced Path Planning for Mobile Entities