A crosscutting application which is driving a significant fraction of the research in both the Coordinated Robotics Lab and the Flow Control Lab is that of the accurate estimation & forecasting of contaminant plumes leveraging sensor-laden unmanned vehicles. The work was originally motivated by the problem of coordinating emergency responses to chemical/radioactive/biological plumes in homeland security settings, such as plumes resulting from possible plant explosions (e.g., Chernobyl) or dirty bombs. This class of problems received renewed international interest in 2010 due to the Gulf-coast underwater oil plumes and the Icelandic ash plumes, and is sponsored by both National instruments and the U.S. Department of Energy.
Motivated by such applications, we have developed a new Hybrid (variational/Kalman) Ensemble Smoother (HEnS) algorithm for state estimation in large-scale systems in the face of substantial nongaussian uncertainties, both in the source strength and location, and in the winds/currents driving the evolution of the plume. This new algorithm effectively combines the principle strengths of the two most effective approaches available today for weather forecasting [ensemble Kalman filtering (EnKF) and space/time variational (4Dvar) methods], while retaining their numerical tractability for large-scale systems. We are also developing a closely related hybrid Targetted Adaptive Observation (TAO) algorithm for coordinating the motion of sensor-laded unmanned vehicles in such systems in a feasible manner (that is, respecting the practical constraints on the motion of the sensor vehicles), again leveraging both ensemble and variational approaches. This algorithm targets as a cost function not the plume itself, but the principle uncertainties of the plume location at the forecast time, and addresses this uncertainty by optimizing feasible trajectories of the sensor-laden vehicles in order to collect the most valuable information possible to minimize this uncertainty.
Both algorithms mentioned above, in addition to being theoretically rigorous, have already proven to be uniquely effective on representative model problems, as reported here. On July 30, 2010, we began testing both of these algorithms, in 2D, using a fleet of a dozen sensor-laden Switchblade vehicles and a heavy smoke plume released in a parking lot near our labs.
A series of 8 tests were performed on July 30, with the last two being comprehensive tests which brought all the pieces together, once the various bugs from the earlier tests were shaken out or circumvented. These tests were covered live by multiple media outlets, including the dynamic coverage performed by
A "eye-in-the-sky" camera hanging from a large balloon overhead did time-lapse photography to record how well the forecasting system performed, as shown below.
[Note: if the movies below don't directly work on your browser, you can download them directly here:
The animation above is a time-lapse movie of an inert smoke plume and the switchblade vehicles moving through it, performed by the eye-in-the-sky camera supported by a large balloon. The vehicles in this particular test were moving in a deliberate fashion under the control of the remote supercomputer, which coordinated their motion so as to be of maximum use for the forecasting problem the supercomputer was working on. This particular test, performed at 8:42am on July 30, 2010, represents the first, complete, sucessful system integration of substantial research components of the PhD theses of Andrew Cavender, Joe Cessna, Chris Colburn, Nick Morozovsky, and David Zhang in an actual environmental flow system in real time. Kim Wright built the frame for the eye-in-the-sky camera frame. Eliot Kroo performed the registration of the video images. Several other MS and undergraduate students were also closely involved.
As more post processing of the test shown above is performed, we will provide further animations, pictures, and video at this web page, so please come back soon! In the meantime, a preliminary animation of the computer forecast of a plume is shown above. The concentration of the plume in the "truth" model shown in the upper left. The "estimate" of the plume concentration, based only on the measurements taken by the sensors (moving dots), is shown in the lower left. The uncertainty of the x component of velocity (top), the y component of velocity (middle), and the plume concentration (bottom) is shown on the right. The tracks in the subfigure in the lower-right corner show the planned upcoming vehicle trajectories - the vehicles execute the first 20% of each plan, then a new plan is made. This framework is known as receding-horizon model predictive control. The plan is optimized by the supercomputer computation in order to send the vehicles on the best trajectories which minimize the model of the forecast uncertainty, while respecting the dynamic constraints on the speed, acceleration, and turning radius of the vehicles.
Note that we are also collaborating with Prof. J. Kosmatka and coworkers at UCSD in order to test these algorithms using multiple UAVs in large airborne plumes in the years to come.