Binary Tracking for Spatial QA and Navigation with Open Vision-Language Models

A team of robotics researchers has proposed BinTrack, an open-source spatial-localization agent designed to help service robots answer location-based queries more accurately and quickly. The system improves accuracy by up to 22.8% and accelerates inference speed by 1.5x over prior open-source methods, according to a paper submitted in 2026 [1]. The work addresses a core challenge for service robots: answering spatial questions like "where can I find a dry cleaner on the way back home?" while navigating long routes [1]. Current approaches often depend on retrieval-augmented agents built on closed-source models such as GPT-4o. However, the researchers note that real-world robots frequently cannot rely on these models due to network instability, communication latency, and deployment costs, creating a need for open-source solutions that can run onboard [1]. BinTrack leverages the temporal ordering of a robot's trajectory to perform a binary search over segments between two anchor landmarks identified from a user's query [1]. The system was evaluated on the SpaceLocQA benchmark, a standard for spatial question answering. It achieved state-of-the-art results, even matching the reported performance of closed-source models on the benchmark's most challenging global category, a setting that previously required strong reasoning agents [1]. Beyond the benchmark, the team released GangnamLoop, a new multi-trip outdoor dataset collected by deploying a real quadruped robot on public streets. The dataset revisits the same locations under different outdoor conditions and pairs the robot's low viewpoint with that of a human owner, providing a practical testbed for future research [1]. The availability of high-quality datasets is a critical driver of progress in machine learning, as advances often stem from the release of well-constructed training and evaluation data alongside algorithmic improvements [2]. The source code and datasets for BinTrack and GangnamLoop have been made publicly available [1]. The paper was submitted to arXiv on 15 June 2026 under the title "Binary Tracking for Spatial QA and Navigation with Open Vision-Language Models" [1].