The performance of the algorithm is greatly affected by the perspective of the collected point cloud

[CaptainWuDaoKou]

I found that the algorithm seemed to only achieve the best performance for point clouds collected from certain viewpoints, and for point clouds collected from other viewpoints or after the coordinate transformation of the point cloud, the results of the grasps generated by the algorithm would be poor or not generate usable grasps at all. Is there an optimal point cloud acquisition perspective? Or is there a way to transform the collected point cloud to this perspective?

Thanks