ARCHER RIGHT ON TARGET
This robot only took eight trials to figure out how to hit the center
of a bullseye. iCub is using a learning algorithm called ARCHER
(Augmented Reward Chained Regression) which is optimized for tasks
that have an easily definable goal and measurable progression towards
that goal. Basically, hitting the center of the target equates to a maximum
reward, and the algorithm builds off of past experience to estimate how
to alter iCub’s hand positions to improve the aim of the arrow. In this
case, the distance between iCub and the target is only 3. 5 meters, but
there’s no reason it couldn’t be scaled up to larger distances.
This robot experiment was conducted by Dr. Petar Kormushev,
Dr. Sylvain Calinon, and Dr. Ryo Saegusa at the Italian Institute of
ARCHER uses a chained local regression process that iteratively estimates new policy parameters which have a greater
probability of leading to the achievement of the goal of the task, based on the experience so far. (Huh?) An advantage of
ARCHER over other learning algorithms is that it makes use of richer feedback information about the result of a rollout.
For the archery training, the ARCHER algorithm is used to modulate and coordinate the motion of the two hands while
an inverse kinematics controller is used for the motion of the arms. After every rollout, the image processing part recognizes
automatically where the arrow hits the target which is then sent as feedback to the ARCHER algorithm. The image recognition
is based on Gaussian Mixture Models for color-based detection of the target and the arrow’s tip. iCub has 53-DOF and
is 104 cm tall.
This research will be presented at the Humanoids 2010 conference December 6-8, 2010, in Nashville, TN.
Ronald Arkin and Alan Wagner
TELL ME SWEET LITTLE LIES
A robot deceives an enemy soldier by creating a false trail and then
hiding so that it will not be caught. While this sounds like a scene
from one of the Terminator movies, it's actually the scenario of an
experiment conducted by researchers at the Georgia Institute of
Technology as part of what is believed to be the first detailed
examination of robot deception.
"We have developed algorithms that allow a robot to
determine whether it should deceive a human or other intelligent
machine, and we have designed techniques that help the robot
select the best deceptive strategy to reduce its chance of being
discovered," explained Ronald Arkin, a Regents professor in the
Georgia Tech School of Interactive Computing.
Can robots lie? This is the subject of the research that enables
them to detect whether one is susceptible and use that gullibility
against another based on algorithms. They feel that the deception
could help it avoid being captured by evil military sources and calm
those they rescue, but robotic experts claim that this ability would
not only damage robot’s current positive image, but may lead to
teaching them to be able to gamble and hunt.
24 SERVO 11.2010