The objective of WP4 is to develop, test, and evaluate efficient, robust and versatile
distributed control methodologies for the sFLY multi-MAV system; as already mentioned in
subsection 1.2.5 the approach adopted within WP4 is to appropriately combine existing
computationally efficient optimization-based DCMR methods (appropriately
modified/extended to meet the needs of the particular sFLY applications) for distributed
control of multi-MAV systems with the new Cognitive-based, Adaptive Optimization (CAO)
methodology of [Kosmatopoulos2007]-[Kosmatopoulos2008b] that will be used for the
online optimization (calibration/adjustment) of the optimization-based DCMR algorithm’s
tuneable parameters.
Within each WP4 task, an integrated software prototype (integrating the particular
optimization-based DCMR methodology to be chosen/designed within each task with the
CAO methodology) will be developed; the development of this software prototype will take
into account the sFLY hardware, communication and software requirements and
constraints as will be defined in WP1-WP3 in order to be delivered in a form deployable
within the sFLY hardware/software architecture. Also, a simulation model of the sFLY
swarm of MAVs performing distributed control tasks will be developed in order to verify the
integrated software prototype operation, test and preliminary evaluate the proposed
methodology and calibrate, if necessary, the software design parameters and procedures.
Further calibration may be also imposed, if deemed necessary, during the implementation
of the software prototype to the sFLY demonstration.
WP4.1: Optimal surveillance coverage with multiple MAVs under
formation and environmental constraints
This task will address the problem of designing efficient distributed control methodologies
for swarms of sFLY MAVs employed in search and rescue operations over remote areas. In
this particular problem, the MAVs trajectories need to be optimized, in real-time, so as a
surveillance coverage criterion (defined as the total surface monitored in a given time
interval) is optimized. Apparently the problem addressed in this task is a dynamic
optimization problem that strongly depends on the terrain morphology and has to take into
account various environmental constraints imposed on the MAVs, such as velocity
constraints, visibility-imposed constraints by occluding objects in case of urban or rough
terrain, etc. The work within this task will follow the following steps:
- As a first step, an overview and evaluation of existing DCMR methodologies for optimal
surveillance coverage using multiple aerial robots will be performed. The approach
that best fits the sFLY constraints and objectives will be chosen and - if deemed
necessary - modified appropriately for the purposes of the particular application
addressed in this task.
- The optimization-based algorithm chosen in the previous step will be combined with
the CAO methodology. The tuneable parameters of the aforementioned optimizationbased
algorithm – to be optimized in real-time by CAO – will be the parameters
defining the MAVs motion model, the robot-environment interaction and the terrain
morphology.
- Finally, an integrated software prototype will be developed, verified, calibrated and
evaluated using the procedure described in the introduction to WP4 above.
WP4.2: Exploration strategies with multiple MAVs under
communication constraints
This task will address the problem where the sFLY swarm is required to operate in
cluttered or dynamically changing environments without external support, in which case
autonomous exploration capabilities are of utmost importance. Exploring autonomously an
unknown environment requires the development of active localization and mapping
algorithms that minimize the uncertainty of the constructed map. This is an optimization
problem whose solution depends on the availability of analytical expressions that describe
the precision of the localization and mapping process as a function of key system
parameters such as the density of landmarks in the area, the number of MAVs involved, the
measurement and communication topology of the MAVs, and the precision of the sensors
used by the MAVs for measuring their ego-motion and sensing the environment. The work
within this task will follow the following steps:
- A gradient descent-based algorithm will be firstly developed for (i) exploring
unmapped terrain and (ii) improving the accuracy of the constructed map by requiring
the coordination of the MAVs so as to concurrently achieve both localization and
mapping in minimum time and with limited communication. This gradient descent
algorithm will be based on the analytical expressions that describe the precision of
the localization and mapping process as a function of key system parameters reported
recently [Roumeliotis2006a], [Roumeliotis2006b].
- Furthermore, and in order to address communication constraints on the MAVs, we
propose to employ optimal sensor-selection strategies for determining and processing
the most informative measurements collected by the MAVs. This problem can be
exactly reformulated to a convex optimization problem [Roumeliotis2006c]-
[Roumeliotis2006e] whose solution provides the frequency that each of the sensors on
the MAVs needs to operate at in order to make optimal use of the communication and
computational resources available to the swarm. In this effort, we will also leverage
our previous work on distributed estimation under severe bandwidth constraints,
where only quantized (1-to-3 bits) observations are communicated between the MAVs.
- The optimization-based algorithms developed in the previous 2 steps, will be
combined with the CAO methodology. The tuneable parameters of the aforementioned
optimization-based algorithms – to be optimized in real-time by CAO – will be the
parameters defining the MAVs motion model, the measurement model and the robotenvironment
interaction.
- Finally, an integrated software prototype will be developed, verified, calibrated and
evaluated using the procedure described in the introduction to WP4 above.
WP4.3: Adaptive estimation of relative position between flying robots
In many applications involving large teams of MAVs, knowing the global position of each
MAV may not be necessary. This is often the case when only the team leader has absolute
localization capabilities while the rest of the team localizes with respect to the leader. In
these situations, it is necessary to design and implement optimal active sensing and
localization algorithms that process robot-to-robot distance (e.g., based on the time-offlight
of communication signals) and/or bearing (e.g., from omni-directional cameras)
measurements and estimate the position of each MAV with respect to the leader. The
challenge in this case, is that there exist numerous singular configurations, especially when
the MAVs move in formation, which reduce the accuracy of the relative localization task
[Roumeliotis2007], [Romeliotis2008a]. The work within this task will follow the following
steps:
- To address the issue of avoiding singular configurations we propose to investigate and
develop optimal motion strategies that trigger small periodic deviations from the
desired formation in order to maximize the acquired relative positioning information
[Roumeliotis2005]. The key parameters that will be considered in this problem are the
number of MAVs, the formation shape and dimensions, the topology of the relative
position measurement graph (RPMG), the sensors’ accuracy, and the existence of
static or moving obstacles along the path of the MAV formation. Each of these
parameters may be time varying and thus the solution to this problem must make
efficient use of the computational resources of the team and adaptively update the
teams motion strategy. For this reason we propose to develop dynamic programmingbased
algorithms that iteratively refine and extend the existing flight plan, while
explicitly considering the parameters’ variability.
- The dynamic programming-based algorithms developed in the previous step will be
combined with the CAO methodology. The tuneable parameters of the aforementioned
dynamic programming-based algorithms algorithms – to be optimized in real-time by
CAO – will be the parameters defining the MAVs motion model, the robot-robot
nonlinear measurement model and the robot-environment interactions.
- Finally, an integrated software prototype will be developed, verified, calibrated and
evaluated using the procedure described in the introduction to WP4 above.
WP4.4: Target tracking under dynamical and communication
constraints This task will address the problem where a team of MAVs is required to follow a moving
object of interest (target) and determine its position and motion over time. Since the
distance and/or bearing measurements of the MAVs are non-linear functions of the sensortarget’s
relative position, the information gain from each observation depends primarily on
the location of each MAV. Minimizing the uncertainty about the target’s position requires
that the MAVs coordinate and determine the trajectory that each of them needs to follow
in order to acquire the most informative measurements. While this problem can be solved
using exhaustive search (one more steps look-ahead) approaches, their complexity
increases exponentially with the number of MAVs involved. Additionally, as it was shown
in [Roumeliotis2008b], motion constraints imposed on the MAVs (e.g., due to the existence
of obstacles or limitations on the MAVs speed) render the optimal tracking problem NP
hard. The work within this task will follow the following steps: - As a first step, an optimization-based algorithm for determining the optimal tracking
strategy of a team of MAVs under dynamic and communication constraints will be
developed. This algorithm will be based on our previous work on distance-based
optimal target tracking [Roumeliotis2008b] by addressing the more general case where
multiple targets need to be tracked using a variety of measurements (distance,
bearing, velocity etc). It is worth noting that in [Roumeliotis2008b] we introduced a
novel relaxation that allowed us to reformulate the problem of minimizing the
covariance (uncertainty) of the target’s position and solve it as a Linear Program.
Additionally, we demonstrated that the resulting optimal motion strategy achieves
accuracy almost identical with that of exhaustive search, with only linear in the
number of sensors cost.
- Furthermore, we will investigate distributed minimization processes, such as
conjugate-gradient [Bertsekas97], which will reduce the computational requirements
per MAV while adhering to the communication limitations of the team.
- Within this step, we intend to combine the approach to be designed in steps 1 and 2
with the CAO methodology in order to address the variability in the system parameters
(e.g., due to the existence of obstacles, multiple targets, and bandwidth availability).
The tuneable parameters of the aforementioned optimization-based algorithms – to be
optimized in real-time by CAO – will be the parameters defining the MAVs motion
model, the nonlinear measurement model and the robot-environment interaction.
- Finally, an integrated software prototype will be developed, verified, calibrated and
evaluated using the procedure described in the introduction to WP4 above.
|