Currently, a typical output is as follows:

 loss: 0.5930 - segmentation_single_pixel_binary_accuracy: 0.6826 - mean_pred_single_pixel: 0.3928 - mean_pred: 0.4465 - mean_true: 0.3928

mean_pred_single_pixel is lower than mean_pred, when it should be the opposite. Based on visual inspection, most of the visible scenes are not positions where a grasp attempt would succeed. This means there is a bias towards successful grasps, because most positions in the world should fail to grasp objects successfully. We need to add additional negative data.

• One way to do this would be to take time steps where the gripper did not close and assume a grasp attempt would fail at those positions, although this would not always be true for every case, it should be for the vast majority.
• Another would be to suggest or collect real poses with delta depth large enough that it wouldn't contact any objects. However, might also be subject to errors.