In
this application, first an IR (infrared) sensor on an airplane flying over
the region is used, and then autonomous vehicles on the ground carrying other
sensors (e.g., GPR and EMI) must move around to improve the discovery and
classification of objects buried under ground.
This
example presents a shift with respect to the traditional paradigm, where
sensor information is used as feedback to the vehicle for control purposes.
As an analogy, consider the difference between using your eyes and senses
to walk from point A to point B and walking around searching for something
that is hidden or difficult to see.
So,
we are developing new decision and control techniques, which allow us to
treat the problems of sensor fusion and inference, as well as control, path
planning and sensor planning in a unified way. The task is to plan robot
navigation and sensor measurements in concert and the objective is to maximize
classification gain and minimize the cost, such as, traveling distance,
time and energy.
A
novel and systematic approach has been proposed to solve this problem,
which belongs to a class of problems, so-called Treasure
Hunt Problem. Using cell decomposition, the pruning algorithm and the
benefit-of-information function, a reduced subset of feasible solutions is
obtained in the form of a pruned connectivity tree, which can be folded
into a decision graph, such as, decision tree or in influence diagram. The
solution of the decision graph is an optimal policy for both robot motions
and sensor measurements. An example of optimal path is shown below. It
achieves much higher path efficiency compared to other methods, such as,
shortest path, complete coverage, random coverage, fixed grid.
q0: robot initial
configuration; qf:
robot final configuration;
blue rectangle:
robot geometry; yellow triangle: sensor field of view;
black: obstacle;
red: target of high information benefit;
magenta: target
of medium information benefit; green: target of low information benefit
Probabilistic
Roadmaps (PRM) can also be utilized to this application, especially for
large-scale landmine fields. An example of optimal path obtained by PRM
method is shown below.
Obstacles: black;
Targets: colored by expected information benefit
Red, High; Green,
Medium; Yellow, Low
Searched path:
black line
Conducting
the research at Duke University
is Dr.
Silvia Ferrari , Assistant Professor in Mechanical Engineering at Duke
University , and her graduate
students, Chenghui Cai
and Guoxian Zhang.
