robotics is the science studying computational
models of robot cognition and its
embedding within robot technologies.
Cognitive robotics merges the science of mental
processes such as attention, perception, learning, short and long term
memorization, reasoning, problem solving, planning and decision making, together
with the science of information processing, sensory feedback and
Cognitive robotics fosters the invention of new
conceived intelligent machines, it does so by integrating a complex body
of components into the design of robot outfitted technologies.
ALCOR LAB research focuses on the following aspects
of Cognitive Robotics:
you can see an example of our Gaze Machine on the left. In the picture I
wear the Gaze Machine mounted on a fire fighter helmet, to protect
myself, as I was going to explore a tunnel where a big accident was
simulated. The Gaze Machine is also available as a pair of glasses, see
here for a description and some videos.
Robot Attention is about what is interesting and why in a scene. What is
interesting is salient. A robot targeting what is salient can exploit high resolution vision in the region of interest, so that all
its processes and computational activities are not overloaded.
reconstruction of large scale outdoor environments:
both our robots
and the gaze machine are endowed with stereo pairs. Robot vision –
likewise the wearable cameras such as the Gaze Machine – cannot be
compared with hand held cameras. A hand held camera incorporates some
implicit attention mechanism, as humans direct instinctively the camera
towards interesting objects and the hand imparts to the camera a smooth
motion which is far different from the robot mounted head or the human
head-eyes system. 3D reconstruction of large scale outdoor environments
requires both localization, and to transform point clouds into meshes to
provide the robot with a good estimate of traversable terrains. 3D
reconstruction is also needed to understand the environment. The
advantage of visual reconstruction, as opposed to laser point cloud, is
that vision can provide a very dense and high resolution reconstruction
where it is needed, because it can be focused. The next step in cognitive robotics is to make
robots able to navigate any kind of environment.
this is studied to make robot proactive. We are now
building a database of motions that can be used for our applications.
Note that human motion is not necessarily the same thing as human action
or human gesture.
robot planning is about a full hierarchy of decision steps, from
elementary control issues, like the velocity necessary to overcome an
obstacle, up to managing tasks and a whole mission.
Decision levels range from early stimuli to the formation of an
appropriate model of the world. A model of the world cannot be defined a
priori nor can be learned abruptly. Understanding the structure of a
good knowledge of what the robot has to do in the world is a major issue
and concerns high level planning, task switching, path planning and
how to manage knowledge acquisition for several robot activities. One
way of talking to your robot and explaining it what should be done under
certain circumstances is to letting the robot experience some specific
event and showing it the right response. This can be done by augmenting
its reality, not yours.