Automated Data Curation Using GPS & NLP to Generate Instruction-Action Pairs for Autonomous Vehicle Vision-Language Navigation Datasets

Guillermo, Roque,; Maquiling, Erika; Tapia, Lopez, Jose Giovanni; Greer, Ross · 2025 · ArXiv.org

DOI: 10.48550/arxiv.2505.03174

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the high cost and inefficiency of manually annotating Instruction-Action (IA) datasets required for training autonomous vehicle (AV) vision-language navigation (VLN) models. The authors propose leveraging Global Positioning System (GPS) navigation applications as an untapped source of natural language instructions. By combining GPS voice commands with video data and vehicle trajectories, the study aims to automatically generate large-scale Vision-Language-Action (VLA) triads without human annotation, thereby accelerating dataset creation for robust AV training. The methodology involves a pilot data collection phase using three common navigation apps: Apple Maps, Google Maps, and Waze. The researchers drove five routes across residential, rural, commercial, and freeway environments in California, collecting 71, 82, and 80 verbal commands, respectively. These commands were manually categorized into eight referentiality classes: Road Names, Distance, Static Objects, Turn, Cardinal direction, Location Name, Lane Information, and Light Information. To demonstrate full automation, the authors developed the ADVLAT-Engine, a prototype system that synchronizes visual perception (video camera), language commands (GPS audio transcribed via OpenAI’s Whisper model), and action streams (GPS logs). This system links specific video frames to verbalized instructions, creating synchronized VLA data. The results characterize the linguistic diversity of GPS instructions, revealing that commands vary significantly in referential cues, such as using distance versus static landmarks. Statistical analysis of the pilot data showed that multi-attribute commands are common, with combinations like "Destination, Road, Turn" appearing most frequently. The study confirmed that the ADVLAT-Engine successfully generates annotated VLA triads by synchronizing video, spatial positioning, and transcribed audio. The automated pipeline effectively replaces manual labor, capturing active decision-making segments while filtering out stationary driving periods. The significance of this work lies in its potential to solve the scarcity and cost issues associated with domain-specific VLA datasets. By automating data curation, the approach enables the rapid generation of high-quality, diverse training data for end-to-end learning models. This method supports contrastive learning approaches and improves the robustness of AV decision-making across different geographic regions. The authors conclude that this automated framework can significantly reduce the performance gap between familiar and new environments, facilitating the development of safer, more interactive autonomous systems.

Key finding

An automated system using GPS app outputs and speech transcription can successfully generate synchronized vision-language-action datasets by categorizing navigation instructions into eight distinct referential classes without manual human annotation.

Methodology

field_study

Sample size: 233

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

StageOutcomeToolModelPromptAttemptsCompleted
discover success author_sweep 2 2026-05-28
archive success canonical_url 1 2026-06-04
extract success cached 3 2026-06-10
clean success clean 1 2026-06-04
chunk success chunk 1 2026-06-04
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-04
enrich success 1 2026-05-28
promote success 1 2026-06-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 2 2026-06-10
tag success vector_similarity 15 2026-06-11
verify success 2 2026-06-10

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).