Nonlinear control and human decision embedding for robotic reconnaissance.

摘要

The use of mobile point sensors for exploring random fields is of interest across many characteristic length scales and in many domains of science and technology. At nanometer scales, various scanning probe microscopy technologies have played increasingly important roles in biology and materials science, and at larger length scales, point-wise sampling of environmental fields has been a significant factor in improved weather prediction, in effective monitoring of air pollutants, and in gathering data on ocean currents and thermal gradients. This dissertation describes the elements of a new theory of efficient reconnaissance of unknown scalar fields defined on planar domains. The presented approach is a hierarchically organized architecture for automated information acquisition, which relies on two levels of feedback control loops. At the lower level, motion primitives guide the mobile sensor to explore the features of the scalar potential field through closing the loop between sensor measurements and sensor motion. Two motion primitives are designed and analyzed, one that follows and maps level contours (contours with constant potential value), and another that steers the sensor to ascend or alternatively descend the potential function's gradient and, as a result, to localize its extremum points.;The designed motion primitives are later utilized in the higher-level feedback loop as the alphabet of a motion description language for potential field reconnaissance. The common defining characteristic of our reconnaissance protocols is their focus on the rapid discovery of an unknown field's key topological features (i.e critical points and critical level sets). This work has led in a natural way to the creation of parsimonious reconnaissance routines that do not rely on any prior knowledge of the environment. The design of topology-guided search protocols uses a mathematical framework that quantifies the relationship between what is discovered and what remains to be discovered. The quantification rests on an information theory inspired model whose properties allow us to treat the search as a problem in optimal information acquisition. A central theme in this approach is that "conservative" and "aggressive" search strategies can be precisely defined, and search decisions regarding "exploration" vs. "exploitation" choices are informed by the rate at which the information metric is changing.;In addition to defining protocols for automated reconnaissance, this dissertation also presents preliminary work on how human decision embedding can be utilized to improve robotic reconnaissance. For this purpose, a computer game has been designed to simulate reconnaissance of unknown fields. Players carry out reconnaissance missions by choosing sequences of motion primitives from the two families. The strategies that emerge from these choices are classified in terms of both the speed with which information is acquired and the fidelity with which the acquired information represents the entire field.