 
| |
        |
Department of Computer Science ResearchMotor Control and Robotics
Machine learning techniques are often used in robotics. Members of ALL
collaborate with members of the Laboratory for Perceptual
Robotics (LPR) in discussions and projects involving artificial
intelligence and control within the domain of motor control and
robotics. The following are brief descriptions of a few of the
research projects which involve members of ALL or have close ties to
the ALL. The papers describing these projects in greater detail can be
found in either the ALL publications list or the LPR
publications list
Humans exploit dynamics---gravity, momentum, elasticity, and so
on---as a regular principle of motor coordination, yet most robot
control techniques cancel dynamics and solve a movement task in a
brute-force way. Our biologically motivated approach involves the
design of a robot motor learning problem, such that skillful movement
emerges from interaction with the environment. We developed a
three-link manipulator that learns to exploit dynamics and to lift
weights far greater than possible given the static torque
constraints. Joint and contact constraints break traditional control
engineering solutions for acrobot-like problems, and three
underpowered, but actuated joints also make typical machine learning
solutions impractical. The approach used in this work makes use of a
hierarchical control structure that incorporates a number of elements
found in human motor learning: imitation, shaping, knowledge of
results (reinforcement), motor programs, and synergies.
In using a robotic hand to grasp an object, many methods assume that
all relevant geometries are known. However, this can be
computationally expensive and does not allow for reapid generalization
to grasping a novel object. Researchers from the Laboratory in
Perceptual Robotics, including ALL member Andrew Fagg, have worked on
two recent projects involving robot grasping. One project was inspired
by how infants handle novel objects that may be used in a goal. By
interacting with them, they incrementaly build a model of the
object. Based on their experience, the infants and toddlers learn to
grasp an object in a way best suited for its use. For example, in
grasping a spoon filled with applesauce, with the goal of eating the
applesauce, the infant may learn to grasp the spoon in an initially
awkward way in order that when she brings the spoon to her mouth, the
final configuration of the grasp is comfortable. Wheeler, Fagg, and
Grupen show that an intelligent agent can use interaction with the
environment to learn stratagies which accommodate the contraints based
on expected future success. Using these ideas to control a robotic
hand, it is observed that robotic hand strategies are similar to those
used by infants and toddlers. Another project investigates how to
best grasp a novel object, which can be treated as an active
sensory-driven problem. Using a control law, the fingers of a hand can
contract on an object. However, multiple control laws can be used to more
robustly grasph the object. Three control laws are combined by
projecting the actions of subordinate control laws into other control
law nullspaces. The resulting composite controller discovers grasps
that are more robust than single controllers.
In simulations investigating reinforcement learning theories, the
process of learning is often accompanied by costly mistakes. While
this is fine in a simulated environment, actual robots learning
on-line cannot afford such mistakes. Having the learning mechanism
work with preprogrammed abstract actions designed to be compatible
with task constraints, rather than primitive actions, enables a robot
to learn a task without making costly mistakes. M. Huber and R. Grupen
use a set of preprogrammed closed loop controllers as actions to make
a four-legged 12 degree of freedom robot ("The Thing") rotate in an
efficient way. One of the fixed controllers sets three of the legs to
form a stable tripod stance while enabling the fourth leg to
move. Other controllers prevent collisions with obstacles and
optimizing posture to allow for better rotation. These controllers run
concurrently. Reinforcement learning algorithms use these controllers
as actions with to goal to rotate the robot. The use of these
controllers enable the learning agent to find the best policy of
moving the legs without allowing the robot to fall, thus achieving
on-line learning with a real robotic system.
|