Active Terrain-Aided Navigation
Abstract
Navigation without external beacon systems has been a long-standing goal within the autonomous systems community. Terrain-aided navigation (TAN) presents an alternative to external localization but requires a prior map. In many cases, the unmanned vehicle (UV) is surveying areas where no map exists and therefore TAN cannot be employed. Active terrain-aided navigation (ATAN) incorporates reinforcement learning (RL) into TAN to reduce the dependence on a prior map. The spatial and temporal measurement uncertainties create the classical problem of exploration versus exploitation: the desire to explore all map areas while exploiting known areas to minimize position error. A dual stochastic optimal estimation problem models the exploration-exploitation dilemma through an information theoretic framework (ITF) that maximizes information gain in real time: a challenge with high computational complexity. The computational cost is reduced using an RL technique called a partially observable Monte Carlo planning process (POMCP). ATAN employs the ITF and POMCP to provide a better, more flexible coverage plan for TAN. The results of this thesis demonstrate the ability of a UV to navigate autonomously without a prior map while minimizing position error, ensuring total coverage, and creating an accurate map within time/energy constraints. Through ATAN, greater levels of autonomy are exhibited, improving upon the TAN framework but also providing a larger offering to the field of robotics.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2020
- Accession Number
- AD1114229
Entities
People
- Darren Kurt
Organizations
- Naval Postgraduate School