Mapping and Insights

  • Explores novel techniques to map and understand a dynamically changing world

    Creating high-quality 3D virtual models of complex physical objects and environments involves repeatedly sensing the environment and fusing sensor measures into a consistent representation. This requires estimation techniques such as simultaneous localisation and mapping (SLAM). Recent techniques incorporate artificial intelligence methods, such as deep learning, with SLAM processing, and provide not only a geometric representation of the scene but also the semantic of the objects and places; insights of great importance for autonomous robots and intelligent systems. Existing solutions rely heavily on several assumptions, such as that most parts of the environment are static, do not change over time, and include considerable structural and textural features that can be easily tracked. Such solutions tend to fail or perform poorly in complex, real-world scenarios where there are significant changes in appearance and structure, highly repetitive patterns, or lack of texture. Also, estimation techniques tend to be computationally expensive, and very few solutions can run efficiently onboard real robots while providing accurate quantification of uncertainty.