VL-LN Bench: Long-Horizon Navigation with Active Dialogs
Published:Dec 26, 2025 19:00
•1 min read
•ArXiv
Analysis
This paper addresses the limitations of existing embodied navigation tasks by introducing a more realistic setting where agents must use active dialog to resolve ambiguity in instructions. The proposed VL-LN benchmark provides a valuable resource for training and evaluating dialog-enabled navigation models, moving beyond simple instruction following and object searching. The focus on long-horizon tasks and the inclusion of an oracle for agent queries are significant advancements.
Key Takeaways
- •Proposes a new task, Interactive Instance Object Navigation (IION), that incorporates active dialog.
- •Introduces the VL-LN benchmark, a large-scale dataset for training and evaluating dialog-enabled navigation models.
- •Demonstrates significant improvements over baselines using the VL-LN benchmark.
- •Addresses the limitations of existing navigation tasks by focusing on ambiguity and long-horizon goals.
Reference
“The paper introduces Interactive Instance Object Navigation (IION) and the Vision Language-Language Navigation (VL-LN) benchmark.”