A reinforcement learning mobile robot obstacle avoidance method based on large language model guidance

By combining scene vectors generated by a large language model with low-dimensional matrix groups constructed by the DWA algorithm, a reinforcement learning state space is formed, which solves the problems of low obstacle avoidance efficiency and high training cost of traditional DWA algorithm in dynamic environments, and realizes efficient obstacle avoidance and path planning in complex environments.

CN122284599APending Publication Date: 2026-06-26NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NANJING UNIV OF POSTS & TELECOMM
Filing Date
2026-03-26
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Traditional DWA algorithms are inefficient at obstacle avoidance in dynamic environments, while DRL methods have high training costs and insufficient generalization ability, making it difficult to achieve good obstacle avoidance results in real and complex scenarios.

Method used

By combining scene vectors generated by a large language model with a low-dimensional matrix group constructed by the DWA algorithm, a state space for reinforcement learning is formed. The generalization ability and robustness of the model are trained and tested using the PPO algorithm.

Benefits of technology

It improves the obstacle avoidance efficiency and generalization ability of mobile robots in dynamic environments, reduces training costs, and enables real-time obstacle avoidance and path planning in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122284599A_ABST
    Figure CN122284599A_ABST
Patent Text Reader

Abstract

This invention relates to the field of obstacle avoidance technology for mobile robots, specifically a reinforcement learning-based obstacle avoidance method for mobile robots guided by a large language model (LLM). The method includes: using a large language model (LLM) to guide the generation of scene vectors reflecting scene features; and constructing a low-dimensional matrix group conforming to robot dynamics constraints using the Dynamic Window (DWA) method, which together with the scene vectors constitutes the state space for reinforcement learning. This structured, low-dimensional state representation significantly reduces the learning burden of the policy network, enabling the policy to quickly distinguish key states and reduce ineffective exploration in the early stages, thereby reducing training costs. Compared to pure reinforcement learning, which relies on a simulation environment, generating different scene vectors through LLM allows for transfer between different scenes, improving the generalization ability of the method.
Need to check novelty before this filing date? Find Prior Art