Computer-automated robot grasp depth estimation
A computer system trains a neural network to predict, for each pixel in an input image, the position that a robot's end effector would reach if a grasp (“poke”) were attempted at that position. Training data consists of images and end effector positions recorded while a robot attempts grasps in a pick-and-place environment. For an automated grasping policy, the approach is self-supervised, as end effector position labels may be recovered through forward kinematics, without human annotation. Although gathering such physical interaction data is expensive, it is necessary for training and routine operation of state of the art manipulation systems. Therefore, the system comes “for free” while collecting data for other tasks (e.g., grasping, pushing, placing). The system achieves significantly lower root mean squared error than traditional structured light sensors and other self-supervised deep learning methods on difficult, industry-scale jumbled bin datasets.
1. A method performed by at least one computer processor executing computer program instructions stored on at least one non-transitory computer-readable medium, the method comprising:
(A) training a neural network, the training comprising:
(A)(1) receiving, as training data:
a plurality of training images I, each representing a three-dimensional scene at a corresponding time when a robot's end effector is at a corresponding stopping point in the three-dimensional scene; and
a plurality of poses of an end effector of a robot, each representing a corresponding pose of the robot's end effector at a corresponding one of the stopping points; and
(A)(2) training the neural network using the plurality of training images I, and the plurality of poses as the training data, to produce a trained neural network;
(B) applying the trained neural network to predict, for each of a plurality of pixels P in an input image, (1) a depth that the robot's end effector would reach in a three-dimensional environment if the robot's end effector moved into the three-dimensional environment to target a projection of the pixel P onto the three-dimensional environment, and (2) an uncertainty of the depth; and
(C) determining when to slow down the robot's end effector based on the depth and the uncertainty of the depth.
2. The method of claim 1 , wherein (A)(1) further comprises receiving, as training data:
depth data representing, for each of the plurality of training images I, depths in the three-dimensional scene represented by the training image; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the depth data, and the plurality of poses as training data, to produce the trained neural network.
3. The method of claim 2 :
wherein the depth data comprises a plurality of aligned depth images D comprising, for each of the plurality of training images I, a corresponding aligned depth image representing a depth map of the three-dimensional scene represented by the training image; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the plurality of aligned depth images D, and the plurality of poses as training data, to produce the trained neural network.
4. The method of claim 2 :
wherein (A)(1) further comprises receiving, as training data, a plurality of grasp successes indicating whether a plurality of attempts by the robot's end effector to grasp an object in the three-dimensional environment were successful; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the depth data, the plurality of poses, and the plurality of grasp successes as training data, to produce the trained neural network.
5. The method of claim 1 :
wherein (A)(1) further comprises receiving, as training data, a plurality of grasp successes indicating whether a plurality of attempts by the robot's end effector to grasp an object in the three-dimensional environment were successful; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the plurality of poses, and the plurality of grasp successes as training data, to produce the trained neural network.
6. The method of claim 3 , further comprising:
(B) before (A), generating the training data, comprising, for each of a plurality of attempts A corresponding to a plurality of target points T in the input image:
(B)(1) controlling the robot's end effector to move into the three-dimensional environment to target a point, in the three dimensional environment, corresponding to the target point T;
(B)(2) identifying a pose of the robot's end effector at a time when the robot's end effector has stopped moving in the three-dimensional environment; and
(B)(3) generating a training image of the three-dimensional environment at the time when the robot's end effector has stopped moving in the three-dimensional environment;
thereby generating the training data, the training data comprising the plurality of training images I, and the plurality of poses of the robot's end effector.
7. The method of claim 6 , wherein (B) further comprises:
(B)(4) generating an aligned depth image of the three-dimensional environment at the time when the robot's end effector has stopped moving in the three-dimensional environment; and
thereby generating the plurality of aligned depth images D.
8. The method of claim 6 , wherein (B)(1) comprises controlling the robot's end effector to move into the three-dimensional environment to target a point, in the three dimensional environment, corresponding to the target point T, and controlling the robot's end effector to attempt to grasp an object at the point in the three-dimensional environment.
9. The method of claim 8 , wherein (B) further comprises:
(B)(5) generating a grasp success indicating whether the attempt to grasp the object at the point in the three-dimensional environment was successful.
10. The method of claim 6 , wherein (B)(1) comprises:
controlling the robot's end effector to move into the three dimensional environment to target a point, corresponding to the target point T; and
identifying a point, in the three dimensional environment, at which the robot's end effector makes contact with an object in the three dimensional environment.
11. The method of claim 10 , wherein (B)(2) comprises identifying a pose of the robot's end effector at the time when the robot's end effector makes contact with the object in the three dimensional environment.
12. The method of claim 6 , wherein (B)(1) comprises:
(B)(1)(a) controlling the robot's end effector to target the point, in the three dimensional environment, corresponding to the target point T, at a first speed;
(B)(1)(b) determining that the robot's end effector is within at most a minimum distance from an object in the three dimensional environment; and
(B)(1)(c) in response to the determination of (B)(1)(b), controlling the robot's end effector to target the point, in the three dimensional environment, corresponding to the target point T, at a second speed, wherein the first speed is greater than the second speed.
13. A system comprising at least one non-transitory computer-readable medium having computer program instructions stored thereon, the computer program instructions being executable by at least one computer processor to perform a method, the method comprising:
(A) training a neural network, the training comprising:
(A)(1) receiving, as training data:
a plurality of training images I, each representing a three-dimensional scene at a corresponding time when a robot's end effector is at a corresponding stopping point in the three-dimensional scene; and
a plurality of poses of an end effector of a robot, each representing a corresponding pose of the robot's end effector at a corresponding one of the stopping points; and
(A)(2) training the neural network using the plurality of training images I, and the plurality of poses as the training data, to produce a trained neural network;
(B) applying the trained neural network to predict, for each of a plurality of pixels P in an input image, (1) a depth that the robot's end effector would reach in a three-dimensional environment if the robot's end effector moved into the three-dimensional environment to target a projection of the pixel P onto the three-dimensional environment, and (2) an uncertainty of the depth; and
(C) determining when to slow down the robot's end effector based on the depth and the uncertainty of the depth.
14. The system of claim 13 , wherein (A)(1) further comprises receiving, as training data:
depth data representing, for each of the plurality of training images I, depths in the three-dimensional scene represented by the training image; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the depth data, and the plurality of poses as training data, to produce the trained neural network.
15. The system of claim 14 :
wherein the depth data comprises a plurality of aligned depth images D comprising, for each of the plurality of training images I, a corresponding aligned depth image representing a depth map of the three-dimensional scene represented by the training image; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the plurality of aligned depth images D, and the plurality of poses as training data, to produce the trained neural network.
16. The system of claim 14 :
wherein (A)(1) further comprises receiving, as training data, a plurality of grasp successes indicating whether a plurality of attempts by the robot's end effector to grasp an object in the three-dimensional environment were successful; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the depth data, the plurality of poses, and the plurality of grasp successes as training data, to produce the trained neural network.
17. The system of claim 13 :
wherein (A)(1) further comprises receiving, as training data, a plurality of grasp successes indicating whether a plurality of attempts by the robot's end effector to grasp an object in the three-dimensional environment were successful; and
wherein (A)(2) comprises training the neural network using the plurality of training images I, the plurality of poses, and the plurality of grasp successes as training data, to produce the trained neural network.
18. The system of claim 15 , wherein the method further comprises:
(B) before (A), generating the training data, comprising, for each of a plurality of attempts A corresponding to a plurality of target points T in the input image:
(B)(1) controlling the robot's end effector to move into the three-dimensional environment to target a point, in the three dimensional environment, corresponding to the target point T;
(B)(2) identifying a pose of the robot's end effector at a time when the robot's end effector has stopped moving in the three-dimensional environment; and
(B)(3) generating a training image of the three-dimensional environment at the time when the robot's end effector has stopped moving in the three-dimensional environment;
thereby generating the training data, the training data comprising the plurality of training images I, and the plurality of poses of the robot's end effector.
19. The system of claim 18 , wherein (B) further comprises:
(B)(4) generating an aligned depth image of the three-dimensional environment at the time when the robot's end effector has stopped moving in the three-dimensional environment; and
thereby generating the plurality of aligned depth images D.
20. The system of claim 18 , wherein (B)(1) comprises controlling the robot's end effector to move into the three-dimensional environment to target a point, in the three dimensional environment, corresponding to the target point T, and controlling the robot's end effector to attempt to grasp an object at the point in the three-dimensional environment.
21. The system of claim 20 , wherein (B) further comprises:
(B)(5) generating a grasp success indicating whether the attempt to grasp the object at the point in the three-dimensional environment was successful.
22. The system of claim 18 , wherein (B)(1) comprises:
controlling the robot's end effector to move into the three dimensional environment to target a point, corresponding to the target point T; and
identifying a point, in the three dimensional environment, at which the robot's end effector makes contact with an object in the three dimensional environment.
23. The system of claim 22 , wherein (B)(2) comprises identifying a pose of the robot's end effector at the time when the robot's end effector makes contact with the object in the three dimensional environment.
24. The system of claim 18 , wherein (B)(1) comprises:
(B)(1)(a) controlling the robot's end effector to target the point, in the three dimensional environment, corresponding to the target point T, at a first speed;
(B)(1)(b) determining that the robot's end effector is within at most a minimum distance from an object in the three dimensional environment; and
(B)(1)(c) in response to the determination of (B)(1)(b), controlling the robot's end effector to target the point, in the three dimensional environment, corresponding to the target point T, at a second speed, wherein the first speed is greater than the second speed.