A Reinforcement Learning Framework for Autonomous Behavior Tree Construction in Door Traversal Tasks

The intelligent decision-making behavior of autonomous systems has developed to adaptive and well-informedly interact with complicated environments but obtaining intelligent decision-making behavior often comes at the cost of difficulty in condition designing and complexity of the system. In this research, we aims to design and implement a decision-making framework that combines Behavior Tree (BT) and Reinforcement Learning (RL), allowing RL to autonomously construct BTs by learning from the environment and feedback signals from the BT. Additionally, we propose a reward function to address hierarchy problems in a door traversal task. The results presented in this paper demonstrate the effectiveness of the proposed framework and reward function in enhancing the performance of the policies and the constructed BTs.