Method and system for object grasping

1. A method comprising:

receiving an image of a scene comprising a plurality of visible objects;

using an object detector, generating a set of keypoints for an exposed portion of a visible object in the scene, wherein each keypoint defines a position within a coordinate frame of the image;

subdividing each set of keypoints into subsets, wherein each subset corresponds to a predetermined subregion of the visible object;

determining a grasp location for each subset based on the positions of the respective keypoints;

selecting, from a resultant set of grasp locations, a candidate grasp location for the visible object based on a grasp probability score associated with the respective subregion; and

facilitating grasp execution based on the candidate grasp location.

2. The method of claim 1 , wherein the subregions are determined when a historical grasp success rate associated with the exposed portion has less than a threshold performance.

3. The method of claim 2 , wherein the historical grasp success rate is associated with every keypoint generated for the exposed portion.

4. The method of claim 1 , wherein the grasp probability score is determined using a trained neural network which is generated from a set of training data which includes subsets of training keypoints corresponding to a respective predetermined subregion and a grasp outcome for a corresponding grasp location.

5. The method of claim 1 , further comprising updating the grasp probability score of the subregion associated with the first candidate object grasp location based on a grasp outcome.

6. The method of claim 1 , further comprising: generating a second set of keypoints for a second visible object in the scene using the object detector, and determining a second candidate object grasp location for the second visible object, wherein the resultant set of grasp locations comprise the candidate grasp location and the second candidate grasp location.

7. The method of claim 1 , wherein the grasp location comprises a central position of the positions of each keypoint corresponding to the respective subregion.

8. The method of claim 7 , wherein the central position is a weighted average calculated based on a detection probability of each keypoint received from the object detector.

9. The method of claim 1 , wherein the coordinate frame of the image is 2D and each grasp location is 2D.

10. The method of claim 9 , wherein the keypoints are generated in 3D by the object detector.

11. The method of claim 9 , wherein the grasp is executed based on an image depth at the candidate grasp location.

12. The method of claim 1 , wherein the subregions are symmetric relative to a principal axis of the object.

13. The method of claim 1 , wherein the set of keypoints defines a convex hull in the coordinate frame of the image, wherein a union of the predetermined subregions defines a grasp area which is smaller than an area of the convex hull.

14. The method of claim 13 , wherein the grasp area surrounds a confined region associated with a grasp probability which is less than a predetermined threshold.

15. The method of claim 14 , wherein the confined region spans at least of: an orifice, a cavity, or an edge.

16. A method comprising:

generating a keypoint set for each object depicted within an image of a scene comprising a plurality of objects;

subdividing each keypoint set into regional subsets, wherein each regional subset corresponds to a predetermined subregion of the respective object;

determining a grasp location for each regional subset based on positions of the respective keypoints;

selecting a grasp location, from a resultant set of grasp locations, based on the associated subregion; and

facilitating grasp execution based on the selected grasp location.

17. The method of claim 16 , wherein the subregions are determined based on an object geometry.

18. The method of claim 17 , wherein the subregions comprise equal segments aligned along a principal axis of the object.

19. The method of claim 17 , wherein the subregions comprise non-concave regions of the object.

20. The method of claim 16 , wherein the subregions are determined when a grasp success rate for grasp locations, determined based on entire keypoint sets, falls below a threshold.

21. The method of claim 16 , further comprising: based on the set of keypoints, determining an occlusion score associated with each regional subset, wherein the grasp location for each regional subset is further determined based on the occlusion score.

22. The method of claim 21 , wherein the occlusion score is a keypoint occlusion score, the method further comprising: determining a keypoint occlusion score for each keypoint of the set, wherein the grasp location for each predetermined subregion is determined based on unoccluded keypoints within the respective regional subset.