CrossMaps: Confidence-Aware Open-Vocabulary Semantic Mapping for Rover Navigation

A new real-time mapping pipeline called CrossMaps allows rovers to build and query spatial maps using natural language, according to a paper submitted to arXiv on June 15, 2026. The system is designed to help unmanned ground vehicles navigate under partial observability by fusing visual data with confidence estimates. [1] The pipeline, presented by researcher Christian Medeiros Adriano, constructs language-queryable maps from RGB-D camera data. CrossMaps integrates multi-scale CLIP embeddings with a confidence-aware fusion mechanism and a dual-memory architecture consisting of Short-Term Memory (STM) and Long-Term Memory (LTM). The STM aggregates noisy visual observations using geometric, semantic, and temporal confidence cues, while cells that reach a threshold of confidence and coherence are promoted to the LTM as persistent semantic landmarks. [1] The system is built for deployment on a Jetson Orin-powered unmanned ground vehicle (UGV) operating alongside a simultaneous localization and mapping (SLAM) framework. It runs in real time and produces semantic heatmaps that an operator can query with natural language to guide navigation. [1] CrossMaps extends prior work in the VLMaps family of approaches, which fuse vision-language model embeddings into metric maps. By adding a confidence layer that tracks sensor quality factors such as range reliability, lighting artifacts, and data density, the pipeline aims to make map updates more robust when a rover has only a partial view of its surroundings. [1] The paper appears amid a broader push to give mobile robots richer semantic understanding of their environments. Open-vocabulary mapping techniques allow robots to recognize and locate objects described in free-form text rather than a fixed set of pre-trained classes, a capability that has drawn attention in planetary exploration and terrestrial logistics alike. [1] Submission records show the manuscript was posted to arXiv on June 15, 2026, under the computer science subcategory of robotics. The work has not yet been peer-reviewed. [1]