Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PSC-AVDN, a training-free aerial vision-and-dialog navigation framework combining LLM instruction parsing, search and confirmation chain-of-thought reasoning with a structured spatial memory, evaluated on ANDH and ANDH-Full.

Problem

Aerial Vision-and-Dialog Navigation (AVDN) for high-altitude UAVs currently relies on supervised finetuning, which is costly; naively applying MLLMs yields unreliable navigation due to weak directional grounding and lack of spatial memory.

Approach

Decouples instruction parsing (LLM converts clock/degree/compass directions to absolute headings), hierarchical search via S-CoT, and fine-grained confirmation via C-CoT, augmented by a Structured Spatial Memory that combines multi-scale views, a trajectory-based visual memory, and an MLLM-generated reference grid map.

Key results

• On ANDH Unseen Val, PSC-AVDN reaches SR 22.6, SPL 17.8, GP 39.2, versus 9.2 SR for the Qwen-VL-Max baseline in the training-free setting.
• On ANDH-Full Unseen Test, PSC-AVDN obtains SR 14.4 and SPL 12.1 (GP 54.5), the best among training-free methods and competitive with supervised methods.
• Ablation shows each stage adds gains: parsing raises SR from 9.2 to 14.6, and the full three-stage pipeline with SSM reaches SR 22.6 on Unseen Val.

Surprising findings

• GPT-4o as a training-free baseline performs very poorly (SR around 2-3%), far below Qwen-VL-Max.

Limitations

• Performance on ANDH-Full Unseen Test and some GP metrics remains below supervised methods such as OpenFly.
• Relies on closed-source MLLM API (Qwen-VL-Max) and multi-step prompting, with a maximum of 3 execution steps.

Experimental setup

Datasets: ANDH ANDH-Full

Benchmarks: ANDH ANDH-Full

Metrics: SR SPL GP

Baselines: E.T. HAA-T LSTM HAA-LSTM TA-GAT FELA OpenFly GPT-4o Qwen-VL-Max

Builds on: Open-Nav AVDN (HAA) MapNav DiscussNav-GPT4

Closest papers by meaning

• SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery (NeurIPS 2026 · similarity 0.88)
• Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation (CVPR 2026 · similarity 0.87)
• Learning Fine-Grained Alignment for Aerial Vision-Dialog Navigation (AAAI 2025 · similarity 0.87)
• WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.87)
• Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology (ICLR 2025 · similarity 0.86)
• Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language Navigation (NeurIPS 2025 · similarity 0.86)