ClutterGen: A Cluttered Scene Generator for Robot Learning

Поделиться
HTML-код
  • Опубликовано: 10 янв 2025

Комментарии •

  • @wuliwuli241
    @wuliwuli241 5 месяцев назад +1

    Using the 3D bounding box of objects for positioning with a bin packing algorithm, or even just using the 2D bounding box information in the xy plane to perform bin packing, could solve this problem. Why design such a complex system to address such a basic issue?

    • @yj-th2hw
      @yj-th2hw 4 месяца назад

      This is a very good question! There are several limitations of bin packing algorithms compared to our methods. First, our task is to determine a physics-compliant stable pose for the queried object with even irregular shape. The main challenge is finding the desired position in a cluttered environment where collisions are sometimes acceptable, such as in stacking, which is not allowed in packing algorithms. You can refer to some generated scene setups at time 2:17. Secondly, our method also considers the diversity of generated scene, which is essential for robot training. However, packing algorithms always place objects in fixed or heuristic ways. Finally, our method can zero-shot generalize to different queried regions after training, while even the most efficient 2D bin packing algorithm still requires O(nlogn). Moreover, our task operates in 3D with object rotation. Let us know if you have further feedback!