WE-BTR: A Behavior Tree Recommendation Method Based on Word Embedding
Hang Su, Li Fu, Xueying Wang, Jinghua Li, Yunlong Wu, Yanzhen Wang · 2023
The widespread application of behavior trees in modeling complex robotic tasks can attribute to their modularity and reusability. Nonetheless, the selection of required behavior trees from previously built databases can be a daunting task, especially when dealing with vast amounts of data. In this paper, we propose a word-embedding-based method for recommending behavior trees (WE-BTR), which focuses on expanding recommendations for candidate behavior trees found through keyword searches. The method segments the complete behavior tree into subtasks and then embeds them as vectors. These vectors are aggregated to generate a vector representation of the behavior tree. By this method, task-relevant information inherent in the behavior tree is precisely captured, enabling an efficient association search and recommendation. Finally, we use cosine similarity to recommend behavior trees similar to the candidate trees. We validated the effectiveness of our method using a sample dataset and demonstrated its superiority in embedding vector, clustering, and recommendation performance.