Translating Natural Language Instructions for Behavioral Robot Indoor Navigation with Attention-History Based Attention
Pengpeng Zhou, Hao He · 2020
In typical indoor robot navigation scene, it is reasonable to use natural language to navigation due to GPS signal lacking. Translating natural navigation instructions to an executable behavior plan can be implemented by a traditional encoder decoder model with attention mechanism. In traditional attention mechanism, the attention distribution is generated based only on the current decoder state, which may ignore previous useful patterns and introduce translation error. In this work, we propose an attention-history reader network that captures the patterns in the attention history. The results test on the Stanford Navigation Datasets show that our method achieves higher scores than the traditional model.