Incorporating Human Domain Knowledge into Large Scale Cost Function Learning

摘要: Recent advances have shown the capability of Fully Convolutional Neural Networks (FCN) to model cost functions for motion planning in context learning driving preferences purely based on demonstration data from human drivers. While pure demonstrations framework Inverse Reinforcement Learning (IRL) is a promising approach, we can benefit well informed priors and incorporate them into process. Our work achieves this by pretraining regress manual function refining it Maximum Entropy Deep Learning. When injecting prior knowledge as network, achieve higher robustness, more visually distinct obstacle boundaries, ability capture instances obstacles that elude models learn data. Furthermore, exploiting these priors, resulting accurately handle corner cases are scarcely seen data, such stairs, slopes, underpasses.