This project is concerned with understanding the management of
visually guided behavior in the face of dynamic environments and
multiple goals. This project has two aims. The first is to address
engineering questions concerning the control of embodied agents. The
second is to address the related question of how humans handle
visuo-motor tasks.
Nathan Sprague (Ph. D. Thesis Work)
Dana Ballard (Thesis Advisor)
Nathan Sprague
The virtual human performing a navigation task. The task is to stay
on the sidewalk while avoiding the blue obstacles and picking up
purple litter. The colored rays show simulated fixations. Blue rays
are sidewalk fixations, red rays are obstacle fixations, and green
rays are pickup fixations.
This project is concerned with understanding the management of
visually guided behavior in the face of dynamic environments and
multiple goals. This project has two aims. The first is to address
engineering questions concerning the control of embodied agents. The
second is to address the related question of how humans handle
visuo-motor tasks.
The research platform is a graphical humanoid that must navigate
through a realistically rendered urban environment. The virtual
human's control architecture is built on the premise that complex
control problems can be handled by sequencing and combining simple
visuo-motor routines that each handle a single well defined task. In
the robotics community this approach, referred to as behavior based
control, has gained wide acceptance. A challenge for behavior based
control is that embodied agents have inherent resource restrictions;
eyes can only look in one direction at a time, and limbs can only be
used for a single task at once. It is an open question how best to
fairly distribute these limited perceptual and physical resources
between concurrently active routines with potentially conflicting
demands.
We address the resource allocation question from a decision theoretic
perspective: resources are distributed preferentially to those
routines that stand to benefit the most. Reinforcement learning
algorithms are used to construct a mapping from action choices to
expected return.
http://www.cs.rochester.edu/~sprague/research.html
virtual human, nathan sprague, dana ballard