New dataset

[ut-austin-rpl]

New dataset

[Steve-Tod]

Hi,

1. The inputs should be the TSDF voxels after fusion.
2. The pos is the coordinate of the point where we want to query the grasp parameters.
3. Occupancy points are randomly sampled inside the volume. The occupancy values are queried from the ground truth mesh, like here. I'm not sure if you really need the ground truth occupancy values if you only want to run the pre-trained GIGA.

[ishikaalunawat]

Hey,

Given I have the TSDF voxels in (40,40,40) format, I would like to know how to:

1. Sample points or generate pos for querying grasp params? According to your dataset and code it seems like there was a grasps.csv from where x, y, z were taken and then scaled to match to voxel size. How did you generate the x, y, z in the csv file?
   Can I randomly sample using np.random.randint(low=0.0,high=40.0,size=(num_query_points,3))*voxel_size ?

[Steve-Tod]

Hi, for grasp data generation, please refer to this file. We follow VGN and randomly sample points on the point cloud observation.

GIGA/scripts/generate_data_parallel.py

Line 76 in d67c438

point, normal = sample_grasp_point(pc, finger_depth)