Parse Search and Confirmation Training Free Aerial Vision and Dialog Navigation with Chain of Thought Reasoning and Structured Spatial Memory

Jul 16, 2026

This paper addresses the Aerial Vision-and-Dialog Navigation (AVDN) task in a training-free setting for resource-efficient high-altitude UAV navigation. Directly applying multimodal large language models can yield unreliable trajectories due to weak directional grounding and missing explicit spatial memory. PSC-AVDN introduces a three-stage Parsing-Search-Confirmation pipeline tightly coupled with Structured Spatial Memory (SSM). Parsing produces stable geometric cues, Search uses chain-of-thought exploration, and Confirmation verifies candidate regions while SSM fuses multi-scale observations, visual memory, and structured geometry for long-horizon consistency.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Abstract

Introduction

Motivation and challenges of high-altitude AVDN

Training-free approach with MLLM baseline and its limitations

PSC-AVDN: Parsing-Search-Confirmation and Structured Spatial Memory

Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory Yu Qi 1 Hongyu Li4 Shaofei Huang5 Tianrui Hui 1 ,2 ,3 * Yaxiong Wang 1 Lechao Cheng 1 Zhun Zhong 1 * Si Liu4 Meng Wang 1 1 School of Computer Science and Information Engineering, Hefei University of Technology 2Jianghuai Advance Technology Center 3Anhui Provincial Key Laboratory of Humanoid Robots 4 School of Artificial Intelligence, Beihang University 5University of Macau arXiv :2607 . 11529v1 [ cs .CV] 13 Jul 2026 Abstract In this paper, we tackle the Aerial Vision-and-Dialog Navigation (AVDN) task in the training-free setting for resourceefficient high-altitude UAV navigation.Naively applying MLLMs leads to unreliable navigation due to weak directional grounding and the lack of explicit spatial memory. To address these issues, we propose PSC-AVDN, a trainingfree framework that tightly couples a three-stage ParsingSearch-Confirmation reasoning pipeline with a Structured Spatial Memory (SSM) . The parsing stage uses an LLM to convert ambiguous dialogue instructions into stable geometric directional and destination cues. A Search Chainof-Thought (S-CoT) then performs stepwise target exploration under high-altitude observations, and a Confirmation Chain-of-Thought (C-CoT) conducts fine-grained verification around candidate regions to resolve visual ambiguity. Meanwhile, SSM integrates three complementary sources of spatial cues, including multi-scale visual observation, spatial visual memory, and structured geometric memory to provide global spatial context and long-horizon consistency. Extensive experiments on ANDH and ANDHFull show that PSC-AVDN establishes new state-of-the-art performance in the training-free setting, matching or surpassing several finetuned methods.Code will be publicly available at: [https://github.com/QY6616/PSC-AVDN](https://github.com/QY6616/PSC-AVDN) 1. Introduction Aerial Vision-and-Dialogue Navigation (AVDN) [7] enables UAVs to follow language instructions while resolving ambiguities through questions and interpreting contextual cues along their flight trajectory. AVDN adopts a highaltitude, top-down perspective similar to remote sensing imagery. Compared with low-altitude urban settings [14, 19, *Corresponding authors. Figure 1 . Motivation of our method. (a) The MLLM baseline suffers from ambiguous directional descriptions and the domain gap between high-altitude imagery and ground-level training data, leading to inaccurate localization. (b) Our PSC-AVDN eliminates directional ambiguity through instruction parsing, performs structured search via chain-of-thought reasoning, and conducts finegrained confirmation around the candidate region to achieve more reliable navigation. In addition, a structured spatial memory is introduced to provide clearer spatial context for reasoning. 37, 40, 45], this high-altitude setting covers a broader visual scope where landmarks are small in scale and can be either densely clustered or sparsely distributed, making it challenging to localize and track fine-grained landmarks during navigation. These characteristics make AVDN well-suited for applications requiring large-scale environmental understanding, such as disaster rescue, environmental monitoring, and geospatial mapping [18, 29, 31, 38], where operators must reason over wide areas relying on visually subtle landmarks. In practice, however, AVDN methods rely on supervised finetuning [7, 35, 36], which entails high computational costs, heavy annotation efforts, and repeated re- annotation and re-training when adapting to new environments. To overcome these issues, we explore a trainingfree AVDN framework that enables resource-efficient UAV navigation from high-altitude perspectives, which is particularly difficult without task-specific training. Thanks to recent advances in Multimodal Large Language Models (MLLMs) [2, 11, 21, 26, 47], we cons

FAQs

How does PSC-AVDN improve reliability compared with the baseline?

PSC-AVDN separates directional understanding from target localization via a Parsing-Search-Confirmation pipeline and adds Structured Spatial Memory that combines multi-scale visual observations, spatial visual memory, and structured geometric memory for better global context.

Jul 16, 2026

Upload by Theodora