I am interested in pathoptimization algorithms which can provide a hard
guarantee on the time required for robots to locate a target with sufficient
accuracy.
For example when
tracking radiotagged carp,
biologists only require GPSlevel accuracy in the estimate of the carp's position, but the fish should be localized as quickly as possible. The radiotag used is pictured to the right.
Single Robot
First, we examine the case of using a single robot, equipped with a directionalsensitive radio antenna to locate a loitering (stationary) radiotagged fish. This closely approximates the human efforts to locate an aggregation of radiotagged fish during the winter months. There are three main difficulties:

Constructing a bearing measurement from radiosignal strength requires many samples. Since the tags used in this domain transmit infrequently, measurements take significant time (12 minutes). Thus, the cost of taking measurements can dominate the total cost to localize.

The antenna is symmetric, producing a bimodal posterior pdf.
An exponential number of terms are required to track the target pdf, even assuming Gaussian noise and priors.

Bearing measurements built from diretionalsensitive antennas are typically very suseptible to outliers (i.e., they have high measurement noise).
My work is on exploiting the mobility of the robot to deal with these problems directly.
First, a mobile robot can move to informative locations while optimizing the trade off between information, travel cost, and measurement time.
It is possible then, to show a closedform bound on the cost to localize a target given system parameters like sensor noise, velocity, measurement time, etc.
Second, by carefully selecting measurement locations, we can ensure that the effects of sensor ambiguity are minimized. For example, in the figure on left, a simple EKF when used with our measurement selection algorithm performs as well as, or better than, other more computationally expensive filters. This shows the effect of correct measurement location choice on the complexity of the posterior pdf.
Finally, we show that no other algorithm can do significantly better than ours by providing a lowerbound on the optimal algorithm, which we show to be close to the performance of our algorithm.
See
[2014 Journal of Field Robotics]
for more information.
MultiRobot Extensions
Second, it is possible to use multiple robots to speed up the localization process by combining their measurements.
However, field robots have a limited communication range that must be considered when choosing measurement locations.
On one hand it may be informative to travel to a wide baseline between measurements, but costly to rendesvous to communicate the results.
Thus, the optimal algorithm must consider communication range in addition to travel time, measurement cost, sensor noise, and the required precision in the final estimate.
In my current work
[Algorithms for MultiRobot Collaborative Localization of Static Targets],
we show lowerbounds on the cost of the optimal manyrobot algorithm, and provide a fieldtested algorithm for choosing measurement communication locations which is provably nearoptimal.
An example of the robot paths from field trials is shown at right (two robots start at the top, and move along the solid and dashed lines respetively). The figure is a satellite image and covers 90 square kilometers. The robots located the tag by taking measurements at the white dots and communicating at the red dots. The final error in the estimate was only 10 meters.
Other work we have done involves developing a mobile, autonomous sensor network [2011 WREM, 2013 IRAM, Systems (below)] and work on scalable multitarget initialization [2013 STAR].
It is also possible to formulate the problem of tracking moving targets as a pursuit / evasion game.
Media
Back