US 12,296,483 B2
Dynamic machine learning systems and methods for identifying pick objects based on incomplete data sets
Shaun Edwards, San Antonio, TX (US); and Daniel Grollman, Boulder, CO (US)
Assigned to Plus One Robotics, Inc., San Antonio, TX (US)
Filed by Plus One Robotics, Inc., San Antonio, TX (US)
Filed on Apr. 12, 2022, as Appl. No. 17/719,108.
Claims priority of provisional application 63/173,568, filed on Apr. 12, 2021.
Prior Publication US 2022/0324098 A1, Oct. 13, 2022
Int. Cl. B25J 9/16 (2006.01); B25J 19/02 (2006.01)
CPC B25J 9/163 (2013.01) [B25J 9/161 (2013.01); B25J 9/1612 (2013.01); B25J 9/1682 (2013.01); B25J 9/1697 (2013.01); B25J 19/023 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer implemented method for training or retraining machine learning models for picking objects in an automated robotic picking environment, the computer implemented method comprising:
obtaining pick request data associated with an automated robotic picking system, the pick request data identifying a pick scenario where intervention was requested, the pick scenario comprising a scenario where an existing pick model being used by the automated robotic picking system was unable to determine pick data that would result in a successful pick;
obtaining intervention data that resulted in a successful pick by the automated robotic picking system, the intervention data comprising human provided pick data that was provided in response to the requested intervention, wherein the intervention data comprises labeled pick request data;
automatically, in response to obtaining the intervention data, augmenting the labeled pick request data, wherein augmenting comprises determining at least one data point in the pick request data that is not associated with a label and assigning a label to the at least one data point in the pick request data that is not associated with a label, wherein each label indicates whether the corresponding data point is associated with a point on an object where the automated robotic picking system can interface the object in order to pick the object or whether the corresponding data point is associated with a point where the automated robotic picking system cannot interface the object in order to pick the object;
training a new pick model using the augmented pick request data, wherein the new pick model is trained to address at least the pick scenario where intervention was requested;
identifying an updated pick model for use by the automated robotic picking system, the identifying comprising testing the new pick model performance.