Autonomous Robots in Unknown Environments: Semantic SLAM and Preference-based Task Planning in Real Time
Abstract: This dissertation proposal aims to enable the robot to perform a set of tasks with a preference order in a priori unknown environment. Such a robot can find widespread applications in many domains, such as household, warehouse, hospital operations, underwater operations, agriculture operations, military assignments, and search and rescue missions. In this dissertation, we leverage our research in semantic SLAM, object search, and preference-guided planning to build a robot system that can address our overarching research problem.
The contributions of this dissertation can be summarized as follows: First, we propose a custom semantic SLAM algorithm. We address the object-level data association problem and propose a map object initialization scheme with a higher success rate. We also designed a novel semantic loop closure method utilizing the object covisibility graph maintained in the map database of semantic SLAM. This loop closure method can distinguish similar scenes and avoid false loop closure. Secondly, we addressed the object search problem, which can considered a simplified version and a milestone of the original research problem. Semantic Prior knowledge is encoded into a Bayesian Network to facilitate the search for the target object. Thirdly, for preference-guided planning, we solve the dilemma between exploitation and exploration by formally appending the exploration task to the original task set with a preference order. Theoretic guarantees for the robot's behavior are proved mathematically. Finally, we conclude this proposal with the work to be done. For a robot modeled as a Markov Decision Process in an unknown environment, we formulate two conditions based on task feasibility and task probability to determine when the current model is outdated and replanning should occur. Implementing and validating the replanning scheme is the remaining work.
Advisor: Professor Jing Xiao
Committee: Professor Nitin Sanket, Professor Carlo Pinciroli, Professor Xinming Huang