Target positioning module training method and device, robot and storage medium

Abstract

A target positioning module training method is applied to the technical field of robots and comprises the following steps: S1, randomly selecting a simulation picture and a real picture; S2, inputtingthe simulation picture into a preset generator, obtaining a synthesized picture through the real picture, wherein the synthesized picture comprises actual position information of the target object; S3, inputting the synthesized picture and the real picture into a preset discriminator; s4, when the discriminator can distinguish the synthesized picture from the real picture, inputting the synthesized picture into a target positioning module to obtain an estimated position of the target object; and S5, updating parameters of the generator and the target positioning module by using a gradient updating algorithm according to the estimated position and the actual position of the target object, and executing the steps S2 and S3 again until the discriminator cannot distinguish the synthesized picture from the real picture in S6, thereby completing training of the target positioning module. The invention further discloses a target positioning module training method and device, a robot and a storage medium, and the method and device can be applied to actual operation of the robot.

Description

technical field [0001] The invention relates to the technical field of robots, in particular to a target positioning module training method, device, robot and storage medium. Background technique [0002] The rise of deep learning technology in recent years has provided a new research direction for the research in the field of robot control. Deep neural networks can automatically extract features from structured data, have good generalization ability, and have great learning potential with parameter expansion, so they are widely used in robot control. However, traditional deep learning techniques need to address the source of a large amount of training data when applied to robot perception modules. Existing large-scale neural networks generally have tens of millions to hundreds of millions of parameters. The optimization of network parameters relies on backpropagation of errors during training and gradient descent to complete, so the model is heavily dependent on a large am...

AI Technical Summary

Problems solved by technology

However, due to the limitations of the physical engine in modeling the real world, and the noise that is difficult to predict and quantify in the real world, it is often difficult to generalize the deep neural network trained on the simulation experiment environment data to the actual task.

Images