A robot following method, system, robot and storage medium

By segmenting the target using the SAM model and generating pseudo-targets, and then fine-tuning the navigation model using the GRPO algorithm, the problem of target loss when the robot is following an occluded scene was solved, achieving continuous target tracking and improved robustness.

CN122431339APending Publication Date: 2026-07-21WUHAN HUANYU ZHIXING TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
WUHAN HUANYU ZHIXING TECH CO LTD
Filing Date
2026-04-09
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

Existing robot following technology is not robust when the target is occluded or disappears, and the following model is difficult to train, resulting in the inability to continue tracking after the target is lost.

Method used

The SAM model is used for pixel-level target segmentation. The target disappearance area is predicted by combining historical motion trajectory, and a pseudo target is generated to replace the real target. The basic navigation model is fine-tuned by the GRPO algorithm to build a target following model.

Benefits of technology

Accurately locate target positions in complex environments, improve the continuity and robustness of target following, reduce data dependence and training costs, and enhance the model's generalization and adaptability in different scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122431339A_ABST
    Figure CN122431339A_ABST
Patent Text Reader

Abstract

The application provides a robot following method, system, robot and storage medium, and the method comprises the following steps: collecting navigation data of a robot and image data of a target to be followed; segmenting the image data of the target to be followed based on a SAM model to obtain a target mask, and predicting a target disappearance area based on the target mask and a historical motion trajectory; performing memory access verification according to the target mask, and storing the target mask and a corresponding current frame of the target to be followed in a memory module after verification; generating a pseudo target to replace the current frame according to the predicted target disappearance area; constructing a navigation network comprising a feature extraction module and a diffusion model, pre-training the navigation network by using the navigation data to obtain a basic navigation model; and fine-tuning the basic navigation model by using the image data of the target to be followed and the memory frame through a GRPO algorithm to obtain a target following model. The application effectively reduces data dependence and training cost, and improves the continuity and robustness of target following.
Need to check novelty before this filing date? Find Prior Art