Approach: We are computing optical
flow fields and processing them (computing time-to-contact values
for flow vectors) to obtain depth maps, which are dense collections
of distances to the objects around the mobile agent, to get the environmental
structure information. The depth maps are then filtered to identify the
obstacles. These obstacle locations and directions serve as the input to
the local navigation algorithms that we are using and working on to improve.
The agent is then supposed to make its way to the designated target location
by carefully avoiding the obstacles with the help of the local navigation
algorithms. We are currently using two different navigation methods:
Human model: Proposed recently by two scientists at Brown
University (William Warren
and Brett
Fajen) after experimenting obstacle avoidance on human subjects and
analysing the collected data sets. Based on controlling the rotational
acceleration of a mobile agent moving with fixed translation speed.
Potential Field Method: Pioneered by Oussama
Khatib. Based on the idea of producing artificial repulsive forces
around the obstacles, and attractive forces around the target, and then
letting the agent find its way naturally by moving inside the force field.
Once we have robust local navigation capability in a mobile agent, high
level modules can then plan very complex tasks for the agent by abstracting
away the low level details, and can use the local navigation module to
handle them (traverse the shorter legs of the long paths) to accomplish
the planned task.
Project MiReLa: Headed by Prof.
Tim Smithers, University of the Basque Country, San Sebastian, Spain
We are also planning to make stable libraries and other code available
through the project home page under the terms of the GNU General Public
Licence, as they get completed and tested.