| In field environments it is
often not possible to provide robot teams with detailed a priori environment
and task models. In such unstructured environments, robots will need to
create a dimensionally accurate three-dimensional geometric model of its
surroundings by performing appropriate sensor actions. However, uncertainties
in robot locations and sensing limitations/occlusions make this difficult.
A new algorithm, based on iterative sensor planning and sensor redundancy,
is proposed to build a geometrically consistent dimensional map of the environment
for mobile robots that have articulated sensors. The aim is to acquire new
information that leads to more detailed and complete knowledge of the environment.
The robot(s) is controlled to maximize geometric knowledge gained of its
environment using an evaluation function based on Shannon’s information
theory. Using the measured and Markovian predictions of the unknown environment,
an information theory based metric is maximized to determine a robotic agent’s
next best view (NBV) of the environment. Data collected at this NBV pose
are fused using a Kalman filter statistical uncertainty model to the measured
environment map. The process continues until the environment mapping process
is complete. The work is unique in the application of information theory
to enhance the performance of environment sensing robot agents. It may be
used by multiple distributed and decentralized sensing agents for efficient
and accurate cooperative environment modeling. The algorithm makes no assumptions
of the environment structure. Hence, it is robust to robot failure since
the environment model being built is not dependent on any single agent frame,
but is set in an absolute reference frame. It accounts for sensing uncertainty,
robot motion uncertainty, environment model uncertainty and other critical
parameters. It allows for regions of higher interest receiving greater attention
by the agents. This algorithm is particularly well suited to unstructured
environments, where sensor uncertainty and occlusions are significant. Simulations
and experiments show the effectiveness of this algorithm. |
