{"id":241,"date":"2025-08-04T10:23:22","date_gmt":"2025-08-04T14:23:22","guid":{"rendered":"https:\/\/sites.bu.edu\/tron\/?p=241"},"modified":"2025-08-04T10:23:22","modified_gmt":"2025-08-04T14:23:22","slug":"sketching-for-efficient-and-robust-top-down-autonomous-navigation","status":"publish","type":"post","link":"https:\/\/sites.bu.edu\/tron\/2025\/08\/04\/sketching-for-efficient-and-robust-top-down-autonomous-navigation\/","title":{"rendered":"Sketching for Efficient and Robust Top-Down Autonomous Navigation"},"content":{"rendered":"

Overview<\/h3>\n
\n
\n
\n

Inspired by the human ability to efficiently and robustly navigate in a diversity of environments, our project aims to develop a novel top-down, low-resource robotic navigation approach in unmapped environments. Comparing state-of-the-art solutions in robotics with their natural counterparts, this project focuses on three opportunities:<\/span><\/p>\n

    \n
  1. Combine machine learning and optimization techniques to sketch a high-level representation of the environment composed of semantically meaningful units.<\/span><\/li>\n
  2. Generate a collection feedback controllers, with rigorous performance guarantees, that can be used to both navigate and improve the representation of the environment.<\/span><\/li>\n
  3. Consider multi-agent settings where robots in a team use collaboration to further reduce sensing requirements, and reuse previous experience to reduce computation requirements.\u00a0<\/span><\/li>\n<\/ol>\n
    \n
    \n
    \n

    BoxMap: High-Level Mapping and Navigation Supported by Machine Learning<\/h3>\n

    BoxMap is a novel high-level mapping method that uses machine learning algorithms to exploit the structure of sensed partial environments, and update a topological map representing semantic entities (rooms and doors) and their relations.<\/p>\n

    BoxMap has two main components: machine learning for extracting and fusing high-level entities from measurements, and graph-based map construction and navigation.<\/p>\n

    Machine Learning for Extracting and Fusing High-Level Entities from Measurements<\/h4>\n

    We developed a new deep learning architecture that combines parametric representations (for semantic entities) with non-parametric partial maps (to fuse measurements). The architecture includes:<\/p>\n