Granted Patent B2

US 12,552,043 · App. 18/566,332 · Granted Feb 17, 2026

Inventors: David T (Beijing, CN); Chao Xue (Beijing, CN)

Assignee: Jingdong Technology Information Technology Co., Ltd.

B25J9/1697B25J9/16B25J9/161B25J9/163B25J9/1664G06V10/80G06T2207/10024G06T2207/10028G06T2207/20084

Abstract

Claims (85)

1 . A method for controlling a robotic arm, comprising:

acquiring a first depth image and a first color image of a target object;

predicting first predicted values and second predicted values of respective candidate actions of the robotic arm based on the first depth image and the first color image, wherein the first predicted value represents a probability of the robotic arm separating the target object from an adjacent object by executing the corresponding candidate action; and the second predicted value represents a probability of the robotic arm successfully grabbing the target object by executing the corresponding candidate action;

determining a target action based on the first predicted values and the second predicted values of the candidate actions; and

controlling the robotic arm to execute the target action;

wherein determining the target action based on the first predicted values and the second predicted values of the candidate actions comprises:

correcting the first predicted values of the candidate actions based on a contour of the target object indicated by the first depth image;

correcting the second predicted values of the candidate actions based on a central position of the target object indicated by the first depth image; and

selecting the target action from the candidate actions based on the corrected first predicted values and the corrected second predicted values.

2 . The method according to claim 1 , wherein predicting the first predicted values and the second predicted values of the candidate actions of the robotic arm based on the first depth image and the first color image comprises:

obtaining a plurality of input depth images by rotating the first depth image by a plurality of preset angles respectively along a preset rotation direction;

obtaining a plurality of input color images by rotating the first color image by the plurality of preset angles respectively along the preset rotation direction;

predicting the first predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into a first prediction network; and

predicting the second predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into a second prediction network.

3 . The method according to claim 1 , further comprising:

acquiring a second depth image of the target object after executing the target action;

determining a first reward value of the target action by using a classification network based on the second depth image and the first depth image, wherein the first reward value is configured to indicate an effective degree of the robotic arm separating the target object from the adjacent object by executing the target action;

determining a second reward value of the target action based on whether the robotic arm successfully grabs the target object; and

training the first prediction network based on the first reward value, and training the second prediction network based on the second reward value.

4 . The method according to claim 3 , wherein determining the first reward value of the target action by using the classification network based on the second depth image and the first depth image comprises:

extracting features of the first depth image and the second depth image respectively by using a feature extraction layer of the classification network;

obtaining a first fusion feature by fusing the feature of the first depth image with the feature of the second depth image; and

obtaining the first reward value of the target action by performing a classification prediction on the first fusion feature by using a classification layer of the classification network.

5 . The method according to claim 2 , wherein predicting the first predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into the first prediction network comprises:

extracting features of the plurality of input depth images and the plurality of input color images by using a feature extraction layer of the first prediction network, and obtaining a plurality of first fusion feature images by fusing the features of the plurality of input depth images with the features of the plurality of input color images respectively;

reversely rotating each of the first fusion feature images along the preset rotation direction; and

obtaining the first predicted values of the candidate actions by performing an action prediction on the first fusion feature images reversely rotated by using a prediction layer of the first prediction network.

6 . The method according to claim 2 , wherein predicting the second predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into the second prediction network comprises:

extracting features of the plurality of input depth images and the plurality of input color images by using a feature extraction layer of the second prediction network, and obtaining a plurality of second fusion feature images by fusing features of the plurality of input depth images with features of the plurality of input color images respectively;

reversely rotating each of the second fusion feature images along the preset rotation direction; and

obtaining the second reward values of the candidate actions by performing an action prediction on the second fusion feature images reversely rotated by using a prediction layer of the second prediction network.

7 . The method according to claim 1 , wherein acquiring the first depth image and the first color image of the target object comprises:

acquiring a captured original depth image and a captured original color image captured;

obtaining three-dimensional coordinate positions corresponding to pixel points in the original depth image and the original color image under a coordinate system of the robotic arm by converting the original depth image and the original color image to the coordinate system of the robotic arm;

obtaining two-dimensional coordinate points corresponding to the three-dimensional coordinate positions respectively by projecting the three-dimensional coordinate positions to a preset two-dimensional plane;

generating the first depth image based on depths of the pixel points corresponding to the two-dimensional coordinate points in the original depth image; and

generating the first color image based on colors of the pixel points corresponding to the two-dimensional coordinate points in the original color image.

8 . The method according to claim 7 , wherein the preset two-dimensional plane is determined based on a minimum working distance along a preset direction under the coordinate system of the robotic arm.

9 . A computer device, comprising a memory, a processor and a computer program stored on the memory and executable by the processor, when executing the computer program, the processor is configured to:

acquire a first depth image and a first color image of a target object;

predict first predicted values and second predicted values of respective candidate actions of the robotic arm based on the first depth image and the first color image, wherein the first predicted value represents a probability of the robotic arm separating the target object from an adjacent object by executing the corresponding candidate action; and the second predicted value represents a probability of the robotic arm successfully grabbing the target object by executing the corresponding candidate action;

determine a target action based on the first predicted values and the second predicted values of the candidate actions; and

control the robotic arm to execute the target action;

wherein the processor is configured to:

correct the first predicted values of the candidate actions based on a contour of the target object indicated by the first depth image;

correct the second predicted values of the candidate actions based on a central position of the target object indicated by the first depth image; and

select the target action from the candidate actions based on the corrected first predicted values and the corrected second predicted values.

10 . A non-transitory computer-readable storage medium with a computer program stored thereon, wherein, when executed by a processor, the computer program is caused to implement;

acquiring a first depth image and a first color image of a target object;

predicting first predicted values and second predicted values of respective candidate actions of the robotic arm based on the first depth image and the first color image, wherein the first predicted value represents a probability of the robotic arm separating the target object from an adjacent object by executing the corresponding candidate action; and the second predicted value represents a probability of the robotic arm successfully grabbing the target object by executing the corresponding candidate action;

determining a target action based on the first predicted values and the second predicted values of the candidate actions; and

controlling the robotic arm to execute the target action;

wherein determining the target action based on the first predicted values and the second predicted values of the candidate actions comprises:

correcting the first predicted values of the candidate actions based on a contour of the target object indicated by the first depth image;

correcting the second predicted values of the candidate actions based on a central position of the target object indicated by the first depth image; and

selecting the target action from the candidate actions based on the corrected first predicted values and the corrected second predicted values.

11 . The computer device according to claim 9 , wherein the processor is configured to:

obtain a plurality of input depth images by rotating the first depth image by a plurality of preset angles respectively along a preset rotation direction;

obtain a plurality of input color images by rotating the first color image by the plurality of preset angles respectively along the preset rotation direction;

predict the first predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into a first prediction network; and

predict the second predicted values of the candidate actions by inputting the plurality of input depth images and the plurality of input color images into a second prediction network.

12 . The computer device according to claim 9 , wherein the processor is configured to:

acquire a second depth image of the target object after executing the target action;

determine a first reward value of the target action by using a classification network based on the second depth image and the first depth image, wherein the first reward value is configured to indicate an effective degree of the robotic arm separating the target object from the adjacent object by executing the target action;

determine a second reward value of the target action based on whether the robotic arm successfully grabs the target object; and

train the first prediction network based on the first reward value, and train the second prediction network based on the second reward value.

13 . The computer device according to claim 12 , wherein the processor is configured to:

extract features of the first depth image and the second depth image respectively by using a feature extraction layer of the classification network;

obtain a first fusion feature by fusing the feature of the first depth image with the feature of the second depth image; and

obtain the first reward value of the target action by performing a classification prediction on the first fusion feature by using a classification layer of the classification network.

14 . The computer device according to claim 11 , wherein the processor is configured to:

extract features of the plurality of input depth images and the plurality of input color images by using a feature extraction layer of the first prediction network, and obtain a plurality of first fusion feature images by fusing the features of the plurality of input depth images with the features of the plurality of input color images respectively;

reversely rotate each of the first fusion feature images along the preset rotation direction; and

obtain the first predicted values of the candidate actions by performing an action prediction on the first fusion feature images reversely rotated by using a prediction layer of the first prediction network.

15 . The computer device according to claim 11 , wherein the processor is configured to:

extract features of the plurality of input depth images and the plurality of input color images by using a feature extraction layer of the second prediction network, and obtain a plurality of second fusion feature images by fusing features of the plurality of input depth images with features of the plurality of input color images respectively;

reversely rotate each of the second fusion feature images along the preset rotation direction; and

obtain the second reward values of the candidate actions by performing an action prediction on the second fusion feature images reversely rotated by using a prediction layer of the second prediction network.

16 . The computer device according to claim 9 , wherein the processor is configured to:

acquire a captured original depth image and a captured original color image captured;

obtain three-dimensional coordinate positions corresponding to pixel points in the original depth image and the original color image under a coordinate system of the robotic arm by converting the original depth image and the original color image to the coordinate system of the robotic arm;

obtain two-dimensional coordinate points corresponding to the three-dimensional coordinate positions respectively by projecting the three-dimensional coordinate positions to a preset two-dimensional plane;

generate the first depth image based on depths of the pixel points corresponding to the two-dimensional coordinate points in the original depth image; and

generate the first color image based on colors of the pixel points corresponding to the two-dimensional coordinate points in the original color image.

17 . The computer device according to claim 16 , wherein the preset two-dimensional plane is determined based on a minimum working distance along a preset direction under the coordinate system of the robotic arm.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (20)