A task-prior guided embodied robot sparse perception and control system

By introducing task prior information and decoupling attention mechanism into the Transformer architecture, the model attention is explicitly guided, which solves the problems of low sample efficiency and overfitting in robot control, realizes fast sparse perception and feature fusion, and improves the efficiency and robustness of embodied robot control.

CN122425752APending Publication Date: 2026-07-21ZHILAI EMBODIED INTELLIGENT TECHNOLOGY (HANGZHOU) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ZHILAI EMBODIED INTELLIGENT TECHNOLOGY (HANGZHOU) CO LTD
Filing Date
2026-05-13
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

Existing robot control methods based on the Transformer architecture suffer from problems such as low sample efficiency, overfitting, lack of explicit prior guidance for attention mechanisms, and difficulty in duty coupling and optimization under conditions of few samples. These problems result in insufficient generalization ability and slow convergence speed of the model in embodied robot control.

Method used

A sparse perception and control system based on task prior guidance is adopted. Through a visual perception module, a task prior encoding module, and an attention guidance and fusion module, task prior information is explicitly introduced to decouple content similarity and prior importance in attention calculation. By using a prior mask generator and a Transformer attention layer, the model’s attention to key regions is directly optimized to achieve sparse perception and feature fusion.

Benefits of technology

It significantly improves the model's learning efficiency and generalization ability under conditions of few samples, shortens the training time, and enhances the model's convergence speed and robustness in the control of embodied robots.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122425752A_ABST
    Figure CN122425752A_ABST
Patent Text Reader

Abstract

The application discloses a kind of embodied robot sparse perception and control system based on task prior guidance, including visual perception module, the visual perception module is used to collect the original perception data of embodied robot work scene, and encode output visual feature label sequence;Control strategy generation module, the control strategy generation module is used to generate robot control instruction sequence according to deep fusion feature decoding;Task prior encoding module, the task prior encoding module is used to obtain the prior information matched with work task, and encode output the prior mask matched with visual feature label sequence.The application can quickly focus on the sparse key information related to task by explicitly introducing task prior guidance, greatly reduces the dependence on massive training data, effectively alleviates the overfitting problem under the condition of few samples.
Need to check novelty before this filing date? Find Prior Art