An Object Placement Optimization System for Efficient and Unbiased Imitation Learning Data Collection

Training data diversity is an important factor in improving the performance of imitation learning. However, object placement diversity and systems that support object placement during data collection have not been sufficiently investigated. In this paper, we analyze the effectiveness of object placement diversity in imitation learning. We evaluate how different placement conditions affect task success using a simulator. Based on the results, we design optimal placement conditions and propose the object placement support system. The proposed system enabled more effective and efficient object placement than human-judged placement, and achieved comparable performance with less data.