Language-Enhanced Mobile Manipulation Framework for Efficient Object Search in Indoor Environments

Liding Zhang, Zeqi Li, Kuanqi Cai, Qian Huang, Zhenshan Bing, Alois Knoll

Keywords: robot manipulation

TL;DR: GODHS framework leverages LLMs to infer scene semantics and guide the robot through a multi level decision hierarchy, mimicking human-like search strategies.

Abstract: Efficient object search in complex environments is critical for applications like household assistance, but traditional methods lack the contextual reasoning for unfamiliar settings. To address this, we propose the Goal-Oriented Dynamically Heuristic-Guided Hierarchical Search (GODHS) framework. GODHS leverages large language models (LLMs) to infer scene semantics and guide the robot through a multi-level decision hierarchy, mimicking human-like search strategies. We ensure the reliability of the LLM's reasoning by using structured prompts at each stage of the hierarchy. For mobile manipulation, we introduce a heuristic-based motion planner combining polar angle sorting and distance prioritization to generate efficient exploration paths. Evaluations in Isaac Sim show that GODHS achieves higher search efficiency compared to conventional, non-semantic strategies. Website and Video are available at: https://drapandiger.github.io/GODHS

Submission Type: Recently published paper within one year (please include the link of the published version in the submission)

Student Paper: Yes

Demo Or Video: No

Public Extended Abstract: Yes

Submission Number: 6