Reinforcement Learning-Based Planar Pose-Graph Optimization

The objective of pose Simultaneous Localization and Mapping (SLAM) or pose-graph optimization (PGO) is to estimate the trajectory of a robot given odometric and loop closing constraints. Traditional state-of-the-art iterative approaches typically involve the linearization of a non-convex objective function and then repeatedly solve a set of normal equations. Furthermore, these methods may converge to a local minima yielding sub-optimal results. In this thesis, the first deep reinforcement learning (DRL) based environment and proposed agents for 2D pose-graph optimization are presented. It is further demonstrated that the pose-graph optimization problem can be modeled as a partially observable Markov Decision Process (POMDP) and performance on real-world and synthetic datasets are evaluated. A proposed agent outperforms second order nonlinear based iterative solvers particularly on challenging instances where it is observed that traditional nonlinear least-squares techniques may fail or converge to unsatisfactory solutions. On the contrary, experiments currently conclude that the learned based approaches presented are still yet inferior to globally certifiable optimizers. Thus, this thesis explores alternative methods to new optimization strategies in the 2D pose SLAM domain.