Question about the variables

[RogerRafa]

Thanks for your great work. However, I am confused about the variables in MinkowskiGraspNetMultifingerType1. Could you please explain the meaning and the difference of the num_multifinger_depth and the num_two_finger_depth?

[longshot-pting]

Thanks for your great work. However, I am confused about the variables in MinkowskiGraspNetMultifingerType1. Could you please explain the meaning and the difference of the num_multifinger_depth and the num_two_finger_depth?

hey, have you get right allegro pose when running this code? I get a (1024,15) shape allegro pose and my code here. Thank you and hope for replying!
`def get_allegro_pose_from_raw_pointcloud(cfgs, pointcloud, extrinsic=np.eye(4), voxel_size=0.005):
"""
输入：
- cfgs: 包含模型路径等参数
- pointcloud: numpy.ndarray of shape (N, 3)，单位为米
- extrinsic: 相机外参（将点云变换到世界坐标），默认为单位矩阵
- voxel_size: 用于稀疏体素化的大小

输出：
- dict: 包含 Allegro 抓取姿态信息（rotation, translation, score, type, etc.）
"""

assert pointcloud.shape[1] == 3, "Point cloud must be of shape (N, 3)"

# 1. 点云变换到世界坐标
points = transform_point_cloud(pointcloud.astype(np.float32), extrinsic)

# 2. 构造稀疏体素张量
coords = np.ascontiguousarray(points / voxel_size, dtype=int)
_, idxs = ME.utils.sparse_quantize(coords, return_index=True)
coords = coords[idxs]
points = points[idxs]
coords_batch, points_batch = ME.utils.sparse_collate([coords], [torch.from_numpy(points)])

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# sinput = ME.SparseTensor(features=points_batch, coordinates=coords_batch, device=device)
sinput = ME.SparseTensor(
    features=points_batch.float(),  # 强制转换为 float32
    coordinates=coords_batch.int(),  # 通常是 int32，这里保险起见显式写出
    device=device
)

# 3. 模型推理
net = get_net(cfgs.checkpoint_path, use_v2=cfgs.use_graspnet_v2)
allegro_models = get_allegro_model(cfgs.allegro_model_path)

end_points = {'sinput': sinput, 'point_clouds': [sinput.F]}
with torch.no_grad():
    end_points = net(end_points)
    grasp_preds, grasp_features = parse_preds(end_points, use_v2=cfgs.use_graspnet_v2)
    # print('grasp preds:',grasp_preds[0])
if len(grasp_preds) == 0:
    print("No grasp predicted.")
    return None

# 4. 后处理
grasp_preds = grasp_preds[0].cpu().numpy()
grasp_features = grasp_features[0].cpu().numpy()
grasp_preds, if_flip = flip_ggarray(grasp_preds)
grasp_features = np.c_[grasp_features, if_flip]

# 排序前 2000
top_indices = grasp_preds[:, 0].argsort()[::-1][:2000]
grasp_preds = grasp_preds[top_indices]
grasp_features = grasp_features[top_indices]

grasp_features_dic = get_graspgroup_features(grasp_features, [sinput])

allegro_depths, allegro_types, scores, grasp_preds, grasp_features = get_allegro_depth_type(
    allegro_models, grasp_features_dic, grasp_preds, grasp_features
)
print('allegro grasp:',grasp_preds[0].shape)
if not RANDOM:
    mask = scores > 0.7
    if not np.any(mask):
        print("❌ No high-score grasp pose.")
        return None
    grasp_preds = grasp_preds[mask]
    grasp_features = grasp_features[mask]
    allegro_depths = allegro_depths[mask]
    allegro_types = allegro_types[mask]
    scores = scores[mask]

grasp_preds, if_flip = flip_z_ggarray(grasp_preds, allegro_types)

# 构造抓取对象
two_fingers_gg = GraspGroup(grasp_preds)
allegro_gg = AllegroGraspGroup()
allegro_gg.from_graspgroup(two_fingers_gg, allegro_types, cfgs.Allegro_mesh_json_path)
allegro_gg.scores = scores
allegro_gg.depths = allegro_gg.depths + allegro_depths + DEFAULT_DEPTH

# 获取得分最高的抓取
idx = np.argsort(scores)[::-1][0]
grasp = allegro_gg[idx]

pose = {
    "rotation_matrix": grasp.rotation_matrix,
    "translation": grasp.translation,
    "score": grasp.score,
    "grasp_type": grasp.grasp_type,
    "depth": grasp.depth,
    "angle": grasp.angle,
}

print("\n✅ Allegro Grasp Pose (From Raw PointCloud):")
print(f"Score     : {pose['score']:.4f}")
print(f"Grasp Type: {pose['grasp_type']}")
print(f"Depth     : {pose['depth']:.4f} m")
print(f"Angle     : {pose['angle']:.4f} rad")
print(f"Translation (x,y,z): {pose['translation']}")
print(f"Rotation Matrix:\n{pose['rotation_matrix']}")

return pose`

[RogerRafa]

Sorry, I only tried using the DH3 to grasp.