US 12,226,904 B2
Robot device, method for the computer-implemented training of a robot control model, and method for controlling a robot device
Anh Vien Ngo, Nehren (DE); Alexander Kuss, Schoenaich (DE); Hanna Ziesche, Leonberg (DE); Miroslav Gabriel, Munich (DE); Philipp Christian Schillinger, Renningen (DE); and Zohar Feldman, Haifa (IL)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Aug. 23, 2022, as Appl. No. 17/893,596.
Claims priority of application No. 10 2021 209 646.8 (DE), filed on Sep. 2, 2021.
Prior Publication US 2023/0063799 A1, Mar. 2, 2023
Int. Cl. B25J 9/16 (2006.01); G06T 7/10 (2017.01)
CPC B25J 9/163 (2013.01) [B25J 9/161 (2013.01); B25J 9/1612 (2013.01); B25J 9/1697 (2013.01); G06T 7/10 (2017.01)] 14 Claims
OG exemplary drawing
 
1. A method for a computer-implemented training of a robot control model, set up to control a robot device for picking up an object of one or a plurality of objects, the method comprising the following steps:
supplying an image, which shows the one or more objects, to a first prediction model of the robot control model to produce a first pickup prediction that includes, for each pixel of the image, a respective first pickup robot configuration vector that describes a configuration of the robot device, with an associated first predicted success probability;
supplying the image to a second prediction model of the robot control model to produce a second pickup prediction, which includes, for each pixel of the image, a respective second pickup robot configuration vector that describes a configuration of the robot device, with an associated second predicted success probability;
supplying the first pickup prediction and the second pickup prediction to a blending model of the robot control model to produce a third pickup prediction, which, for each pixel of the image:
includes a third pickup robot configuration vector that is a combination, weighted by first weighting factors, of the first pickup robot configuration vector and the second pickup robot configuration vector, and
includes a third predicted success probability that is a combination, weighted by second weighting factors, of the first predicted success probability and the second predicted success probability; and
training the robot control model by adapting the first weighting factors and the second weighting factors based on target data in which a successful picking up is assigned to at least one pixel of the image, in such a way that the third predicted success probability produced for the at least one pixel is increased.