Abstract
Sensor management refers to the control of the degrees of freedom in a sensing system. The objective of sensor management is to improve performance e.g. by obtaining more accurate information or by achieving other operational goals. Sensor management is viewed as a sequential decision making process, where decisions at any time are made conditional on the past decisions and measurement data. At the time of deciding a control action for a sensing system the measurement data that will be obtained are unknown. Thus, informally speaking, a solution to a sensor management problem is a policy that determines which sensing action to undertake given the current information on the state of the process under investigation and contingent on any possible realisation of future measurement data outcomes.
This thesis studies sensor management framing the contingent planning problem in the partially observable Markov decision process (POMDP) framework. In particular, applications in mobile robotics are considered. Mobile robots are viewed as controllable sensor platforms.
Based on earlier work on POMDP based robot control, and distinguishing between the two cases of either exploiting or gathering information, we define four canonical sensor management problem types in mobile robotics. In each of the problem types, we exploit the structural properties of their inputs to improve efficiency of applicable contingent planning algorithms.
In particular, we consider sensor management problems for information gathering where the utility of the possible control policies is quantified by mutual information (MI). We identify the relationship between the POMDP formulation of an environment monitoring problem and another contingent planning problem known as a multi-armed bandit (MAB). In a robotic exploration task, we derive a novel approximation for MI.
Through both simulation and real-world experiments in mobile robotics domains, we determine the applicability, advantages, and disadvantages of a POMDP based approach to sensor management in mobile robotics.
This thesis studies sensor management framing the contingent planning problem in the partially observable Markov decision process (POMDP) framework. In particular, applications in mobile robotics are considered. Mobile robots are viewed as controllable sensor platforms.
Based on earlier work on POMDP based robot control, and distinguishing between the two cases of either exploiting or gathering information, we define four canonical sensor management problem types in mobile robotics. In each of the problem types, we exploit the structural properties of their inputs to improve efficiency of applicable contingent planning algorithms.
In particular, we consider sensor management problems for information gathering where the utility of the possible control policies is quantified by mutual information (MI). We identify the relationship between the POMDP formulation of an environment monitoring problem and another contingent planning problem known as a multi-armed bandit (MAB). In a robotic exploration task, we derive a novel approximation for MI.
Through both simulation and real-world experiments in mobile robotics domains, we determine the applicability, advantages, and disadvantages of a POMDP based approach to sensor management in mobile robotics.
| Original language | English |
|---|---|
| Publisher | Tampere University of Technology |
| Number of pages | 172 |
| ISBN (Electronic) | 978-952-15-3701-1 |
| ISBN (Print) | 978-952-15-3676-2 |
| Publication status | Published - 2016 |
| Publication type | G4 Doctoral dissertation (monograph) |
Publication series
| Name | Tampere University of Technology. Publication |
|---|---|
| Publisher | Tampere University of Technology |
| Volume | 1366 |
| ISSN (Print) | 1459-2045 |