Abstract: Mobile robots need accurate, high fidelity models of their operating environments in order to complete their tasks safely and efficiently. Generating these models is most often done via Simultaneous Localization and Mapping (SLAM), a paradigm where the robot alternatively estimates the most up-to-date model of the environment and its position relative to this model as it acquires new information from its sensors over time. Because robots operate in many different environments with different compute, memory, sensing, and form constraints, the nature and quality of information available to individual instances of different SLAM systems varies substantially. `One-size-fits-all` solutions are thus exceedingly difficult to engineer, and highly specialized systems, which represent the state-of-the-art for most types of deployments, are not robust to operating conditions in which their assumptions are not met. This proposal seeks to investigate an alternative approach to these robustness and universality problems by incorporating existing SLAM solutions within a larger framework supported by planning and learning.
The central idea is to combine learned models that estimate SLAM algorithm performance under a variety of sensory conditions, in this case neural networks, with planners designed for planning under uncertainty and partial observability, in this case partially observable Markov decision problems (POMDPs). Models of existing SLAM algorithms can be learned, and these models can then be used online to estimate the performance of a range of solutions to the SLAM problem at hand. The POMDP policy then selects the appropriate algorithm, given the estimated performance, cost of switching methods, and other information. This general approach may also be applicable to many other robotics problems that rely on data-fusion, such as grasp planning, motion planning, or object identification.
Co-advisors: Rod Grupen & Shlomo Zilberstein