STELLA: Smart Trusted Environmental Learning Logic Agent

Abstract

Blind and low-vision (BLV) individuals face challenges in navigating environments and accessing contextual information. Existing assistive systems rely on human guidance or predefined maps, which limits independence. In addition, many approaches focus on single tasks without supporting continuous interaction. In this paper, we propose STELLA (Smart Trusted Environmental Learning Logic Agent), a comprehensive assistive robotic system implemented on a quadruped robot platform that integrates navigation, environment understanding, and conversational interaction. STELLA combines Nav2-based navigation with VLM-based semantic scene understanding. The system is built on multi-agent architecture, where a large language model (LLM) routes user requests to functional modules. It uses a VLM for scene understanding and object search, while Nav2 handles motion-level navigation. User input is captured via speech-to-text (STT), and responses are delivered through text-to-speech (TTS), enabling real-time and private voice interaction. We introduce a hybrid perception strategy that combines VLM-based understanding with object detection for emergency-aware objects. Temporal visual context enables continuous question-answering, allowing coherent follow-up interactions. We demonstrate that STELLA provides a flexible and interactive assistive experience compared to existing approaches.