Granted Patent B2

US 12,397,415 · App. 18/192,420 · Granted Aug 26, 2025

Inventors: Rodrigo Chau (Karlsruhe, DE); Hanna Ziesche (Leonberg, DE); Anh Vien Ngo (Nehren, DE); Gerhard Neumann (Karlsruhe, DE)

Assignee: ROBERT BOSCH GMBH

B25J9/02G06V10/82G06V20/50

Abstract

Claims (22)

1. A method for controlling a robot device, comprising the following steps:

acquiring an image of an environment of the robot device, the environment being a working range of a robot arm of the robot device;

processing the image using a neural network, the neural network being configured to output from the image for at least one action parameter value, a respective value image with pixel values for multiple pixels, the pixel value for each pixel of the multiple pixels indicating an evaluation of an action that is specified by the action parameter value and the position of the pixel in the value image, and the neural network being configured to ascertain an evaluation of each action from an intermediate evaluation of the action and an evaluation of a state of the environment of the robot device, wherein the neural network has an encoder network and a decoder network with a bottleneck between the encoder network and the decoder network and skip connections between the encoder network and the decoder network;

selecting, from multiple actions that characterize a skill of the robot device in each case, a particular action from among the multiple actions for which the pixel value of the pixel in the value image is at a maximum for the action parameter value at the position that specifies the action together with the action parameter value; and

controlling the robot device to carry out the selected action.

2. The method as recited in claim 1 , further comprising:

training the neural network using target values for the evaluations, the target values being ascertained with using a second instance of the neural network according to Double-Deep Q-learning.

3. The method as recited in claim 1 , further comprising:

training the neural network using target values for the evaluations, the target values being ascertained from rewards of control cycles, and the training being carried out in such a way that a sample of the control cycles is given priority over another sample from the control cycles when the evaluation estimated for the sample by the neural network deviates more strongly from the target value for the evaluation than for the other sample.

4. The method as recited in claim 1 , wherein an action that is specified by the position of a pixel is an action that is to be carried out at a position of the environment of the robot device that corresponds to the position of the pixel in the value image.

5. The method as recited in claim 1 , wherein the neural network is configured to output an output which includes a respective value image for multiple action parameter values, each of the multiple action parameter values describing a control variable of an action or a type of an action.

6. The method as recited in claim 1 , wherein the environment of the robot device has multiple objects, and the action parameter values: (i) indicate whether an object is to be displaced or rotated, and/or (ii) indicate directions and/or displacement widths and/or angles of rotation.

7. A robot device including a control device including a processor, the control device configured to control the robot device, the control device configured to:

acquire an image of an environment of the robot device, the environment being a working range of a robot arm of the robot device;

process the image using a neural network, the neural network being configured to output from the image for at least one action parameter value, a respective value image with pixel values for multiple pixels, the pixel value for each pixel of the multiple pixels indicating an evaluation of an action that is specified by the action parameter value and the position of the pixel in the value image, and the neural network being configured to ascertain an evaluation of each action from an intermediate evaluation of the action and an evaluation of a state of the environment of the robot device, wherein the neural network has an encoder network and a decoder network with a bottleneck between the encoder network and the decoder network and skip connections between the encoder network and the decoder network;

select, from multiple actions that characterize a skill of the robot device in each case, a particular action from among the multiple actions for which the pixel value of the pixel in the value image is at a maximum for the action parameter value at the position that specifies the action together with the action parameter value; and

control the robot device to carry out the selected action.

8. A computer-readable medium on which is stored commands for controlling a robot device, the commands, when executed by a processor, causing the processor to perform the following steps:

acquiring an image of an environment of the robot device, the environment being a working range of a robot arm of the robot device;

processing the image using a neural network, the neural network being configured to output from the image for at least one action parameter value, a respective value image with pixel values for multiple pixels, the pixel value for each pixel of the multiple pixels indicating an evaluation of an action that is specified by the action parameter value and the position of the pixel in the value image, and the neural network being configured to ascertain an evaluation of each action from an intermediate evaluation of the action and an evaluation of a state of the environment of the robot device, wherein the neural network has an encoder network and a decoder network with a bottleneck between the encoder network and the decoder network and skip connections between the encoder network and the decoder network;

selecting, from multiple actions that characterize a skill of the robot device in each case, a particular action from among the multiple actions for which the pixel value of the pixel in the value image is at a maximum for the action parameter value at the position that specifies the action together with the action parameter value; and

controlling the robot device to carry out the selected action.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (16)