Abstract
Identifying the origin of hazardous leaks in industrial environments is a critical safety task often managed by autonomous robots. However, standard navigation algorithms frequently encounter local optima in complex, maze-like settings. While human-in-the-loop crowdsourcing has mitigated these failures, it introduces significant latency and cost. This paper presents AutoS2earch, a zero-shot framework that replaces human crowdsourced workers with Multi-modal Large Language Models (MLLMs). By abstracting visual maps into structured linguistic descriptions and employing Chain-of-Thought reasoning, AutoS2earch achieves a 97% success rate, effectively matching human performance while providing a scalable, low-latency solution for emergency response.
1. Introduction
In industrial safety, "source search" involves locating the origin of hazards like toxic gas leaks or radiation. Autonomous robots typically utilize algorithms such as Infotaxis to navigate toward these sources based on sensor data. Despite their utility, these algorithms often struggle in high-obstacle environments, becoming trapped in deadlocks or local optima.
Previous solutions integrated human-AI collaboration, where robots "call for help" via web-based crowdsourcing platforms. However, the reliance on human intervention is problematic during time-sensitive emergencies due to recruitment delays and high operational costs. AutoS2earch addresses this by automating the "rescue" process using Large Language Models (LLMs) capable of human-like spatial reasoning.
2. Methodology: The AutoS2earch Framework
AutoS2earch is a zero-shot framework integrated into a web-based management system. It automates the decision-making process when a robot encounters a navigation failure.
The workflow consists of four primary stages:
- Problem Detection: A monitoring mechanism identifies when a robot ceases progress and triggers a rescue task.
- Visual-Language Conversion: To avoid the spatial inaccuracies common in raw image processing by MLLMs, the system uses an MLLM (e.g., GPT-4o) to interpret the visual map. The model identifies four potential directions (A, B, C, D) and describes them in terms of proximity to the estimated source (represented by green particles) and exploration potential (density of unexplored "black" areas).
- Chain-of-Thought (CoT) Reasoning: A text-based LLM processes these descriptions using a prioritized rule set: ignoring non-existent regions, prioritizing proximity to the source, and then prioritizing unexplored space.
- Execution: The LLM selects a direction, and the robot resumes its search.
3. Experimental Evaluation
The framework was tested on 20 benchmark scenarios featuring a 20m x 20m grid with a 75% obstacle density. These conditions simulate complex industrial layouts where standard algorithms typically fail. The study evaluated several models, including GPT-4o, GLM-4v-plus, and Qwen-VL-plus, comparing their performance against both pure Infotaxis and human-aided teams.
4. Results and Discussion
AutoS2earch demonstrated high effectiveness, achieving a 97% success rate across benchmarks, significantly outperforming the 78.5% success rate of the standalone Infotaxis algorithm. Notably, the Alibaba-developed Qwen model reached a 98% success rate.
The system's efficiency was comparable to human-AI teams in terms of steps taken, but it offered a significant advantage in speed. The average reasoning time was approximately 26 seconds, far faster than the time required to recruit and receive a response from a human worker.
A key insight from the study is the value of linguistic abstraction; LLMs performed better when visual data was translated into structured text, bypassing current limitations in precise coordinate geometry. Furthermore, LLMs were found to follow reasoning rules more strictly than humans, who occasionally exhibited "fuzzy" logic.
5. Conclusion
AutoS2earch successfully demonstrates that MLLMs can replace human crowdsourced workers in complex robotic navigation tasks. While the current study is limited to 2D static environments, it establishes a foundation for zero-shot, scalable human-like reasoning in industrial robotics. Future work will focus on extending the framework to 3D dynamic environments and integrating domain-specific knowledge, such as fluid dynamics, to further enhance search efficiency.
Summary updated