Single-robot indoor structured environment exploration method based on deep reinforcement learning

By employing boundary feature enhancement and non-uniform graph representation methods, the problems of dimensional explosion and boundary detail fidelity in autonomous exploration of large-scale indoor environments are solved, achieving efficient and complete environmental coverage and decision continuity, and improving the robot's autonomous exploration performance in complex environments.

CN122149516APending Publication Date: 2026-06-05NANKAI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NANKAI UNIV
Filing Date
2026-03-11
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing technologies present a contradiction between dimensionality reduction for large-scale environmental inputs and fidelity preservation of key boundary details. Uniform dimensionality reduction schemes are prone to losing boundary topological details, resulting in insufficient exploration coverage and decision stagnation. On the other hand, full-scale fine-grained input schemes face dimensionality explosion and excessive computational overhead, making them unsuitable for the autonomous exploration needs of large-scale indoor structured environments.

Method used

We employ a method based on boundary feature enhancement and non-uniform graph representation. By incrementally constructing a two-dimensional occupied grid map, we identify and label boundary sampling points, construct a non-uniform heterogeneous topological feature map, and combine an attention mechanism with a deep reinforcement learning decision model based on the SAC algorithm to optimize the exploration path.

Benefits of technology

It significantly improves the exploration coverage and success rate in complex structural environments, enhances the continuity of decision-making, reduces repeated paths and decision stagnation, achieves efficient allocation of computing resources, and improves the convergence speed and generalization ability of reinforcement learning models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122149516A_ABST
    Figure CN122149516A_ABST
Patent Text Reader

Abstract

The application discloses a single-robot indoor structured environment exploration method based on deep reinforcement learning, and steps are as follows: a robot acquires environment perception data and self pose data in real time, and completes data initialization; a two-dimensional occupancy grid map is incrementally constructed, and semantic division of an environment region is completed; uniform topological sampling is performed in an explored region, boundary sampling points are synchronously identified and marked; based on a shortest path algorithm, a path skeleton of an exploration map is constructed, non-uniform graph pruning and redundant node elimination are completed; for a region where the boundary sampling points are located, boundary feature dimension expansion processing is performed, and a non-uniform heterogeneous topological feature map is constructed; a deep reinforcement learning decision model based on an attention mechanism and a SAC algorithm is constructed, the non-uniform heterogeneous topological feature map is taken as model input, and an optimal exploration target point of the robot in the next step is output; the robot completes path planning and motion execution according to the output optimal exploration target point, and the circulation is executed until an environment exploration task is completed.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Robot autonomous exploration mapping method, equipment and storage medium

    CN110806211A

  • Mobile robot autonomous exploration method based on deep reinforcement learning

    CN119200601A