Machine learning methods and apparatus for robotic manipulation and that utilize multi-task domain adaptation
Implementations are directed to training a machine learning model that, once trained, is used in performance of robotic grasping and/or other manipulation task(s) by a robot. The model can be trained using simulated training examples that are based on simulated data that is based on simulated robot(s) attempting simulated manipulations of various simulated objects. Portion(s) of the model can also be trained based on real training examples that are based on data from real-world physical robots attempting manipulations of various objects. The simulated training examples can be utilized to train the model to predict an output that can be utilized in a particular task—and the real training examples used to adapt at least a portion of the model to the real-world domain can be tailored to a distinct task. In some implementations, domain-adversarial similarity losses are determined during training, and utilized to regularize at least portion(s) of the model.
1. A method implemented by one or more processors of a robot, comprising:
identifying an image captured by a vision component associated with the robot;
identifying a target object captured by the image;
generating a target object input that denotes position of the target object in the image;
generating a candidate end effector motion vector defining motion to move a grasping end effector of the robot from a current pose to an additional pose;
identifying a current image captured by the vision component, the current image capturing the grasping end effector and at least the target object;
processing the current image, the candidate end effector motion vector, and the target object input using a trained instance grasping model;
generating, based on the processing, a prediction of successful grasp of the target object with application of the motion defined by the candidate end effector motion vector;
generating an end effector command based on the prediction of successful grasp of the target object; and
providing the end effector command to one or more actuators of the robot.
2. The method of claim 1 , wherein the target object input is a segmentation mask that is at least a partial segmentation of the target object in the image.
3. The method of claim 2 , wherein the segmentation mask is determined based on the image.
4. The method of claim 2 , wherein the image comprises multiple channels and a given quantity of pixels in each of the channels, and wherein the segmentation mask has the same quantity of pixels as the image, but includes only a single channel.
5. The method of claim 4 , wherein the position of the target object in the image is denoted, in the segmentation mask, by a subset of pixels, in the single channel, that all have a first value that varies from a second value of all other pixels, in the single channel, that are not in the subset of pixels.
6. The method of claim 5 , wherein the subset of pixels correspond to a predicted location of where the target object is present in the image.
7. The method of claim 5 , wherein the subset of pixels correspond to one or more predicted locations of where the target object is present in the image.
8. The method of claim 5 , wherein the subset of pixels correspond to a bounding box that surrounds the predicted locations of where the target object is present in the image.
9. The method of claim 5 , wherein the subset of pixels consist of a center pixel that is centered on the target object in the image.
10. The method of claim 1 , wherein the target object input is a bounding box that encompasses the target object in the image.
11. A robot, comprising:
an end effector;
actuators;
a vision component;
memory;
one or more processors executing instructions stored in the memory to:
identify an image captured by the vision component;
identify a target object captured by the image;
generate a target object input that denotes position of the target object in the image;
generate a candidate end effector motion vector defining motion to move the end effector from the current pose to an additional pose;
identify a current image captured by the vision component, the current image capturing the grasping end effector and at least the target object;
process the current image, the candidate end effector motion vector, and the target object input using a trained instance grasping model;
generate, based on the processing, a prediction of successful grasp of the target object with application of the motion defined by the candidate end effector motion vector;
generate an end effector command based on the prediction of successful grasp of the target object; and
provide the end effector command to one or more of the actuators of the robot.
12. The robot of claim 11 , wherein the target object input is a segmentation mask that is at least a partial segmentation of the target object in the image.
13. The robot of claim 12 , wherein the segmentation mask is determined based on the image.
14. The robot of claim 12 , wherein the image comprises multiple channels and a given quantity of pixels in each of the channels, and wherein the segmentation mask has the same quantity of pixels as the image, but includes only a single channel.
15. The robot of claim 14 , wherein the position of the target object in the image is denoted, in the segmentation mask, by a subset of pixels, in the single channel, that all have a first value that varies from a second value of all other pixels, in the single channel, that are not in the subset of pixels.
16. The robot of claim 15 , wherein the subset of pixels correspond to a predicted location of where the target object is present in the image.
17. The robot of claim 15 , wherein the subset of pixels correspond to one or more predicted locations of where the target object is present in the image.
18. The robot of claim 15 , wherein the subset of pixels correspond to a bounding box that surrounds the predicted locations of where the target object is present in the image.
19. The robot of claim 15 , wherein the subset of pixels consist of a center pixel that is centered on the target object in the image.
20. The robot of claim 11 , wherein the target object input is a bounding box that encompasses the target object in the image.