Spatio-temporal object persistence modeling and semantics for long-term robot navigation



Journal Title

Journal ISSN

Volume Title



Mobile robots increasingly operate in real-world environments that are subject to change over time. Robots that maintain up-to-date, accurate representations of their environment can more robustly perform long-term autonomous navigation and planning tasks. Uninterrupted robot autonomy is important for many tasks where human intervention is undesirable or impossible including space operations, hazardous environment surveillance, military reconnaissance, and more. This dissertation explores how a robot can adapt maps over long periods of time and predict changes in their environments to maintain long-term navigational autonomy. Spatio-temporal Object Persistence (STOP) models enable a robot to assign temporal characteristics to recognized objects based on their observed persistence in the world. A robot develops these identifying characteristics by iteratively estimating parameters for a Weibull distribution survival model using recursive Bayesian Survival Analysis with Markov Chain Monte Carlo methods. The parameters of the model provide temporally identifying features for semantic classification of objects. These characteristics then allow a robot to estimate the true lifespan of objects and predict when they will leave the environment. A robot updates its belief map based on the modelled temporal behavior of objects in its environment and then perform more intelligent planning and navigation for future operations. Furthermore, once the robot develops temporal class characteristics, it can transfer and apply these characteristics to objects in dynamically similar environments, thus allowing the robot to adapt more quickly to new operational spaces. In this dissertation, we establish the efficacy of STOP models for temporal semantic object classification and show how a robot uses them to adapt a map over time in a semi-static environment. We provide a series of experiments to demonstrate various aspects of our implementation. Our results confirm that the learned models not only predict the temporal behavior of objects in the world but also transfer to unknown, but temporally similar operation spaces where they improve the prediction process.


LCSH Subject Headings