Transcript Robots - BCCN 2009

Real Artificial Life: Robots
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
1
Biorobotics
Question: How can the study of robots help us better understand
biological organisms?
Biorobotics: close collaboration of biologists and roboticists mostly trying
to understand the principles of various forms of locomotion
Biology can ask: “how does phenomenon X work in species Y?”
“analytic approach”
E.g.: How are which parts of the cockroach nervous system controlling its
walking behavior? Experimental limitations: can destroy parts of the system
but not add or replace parts. Can not ask: “does it have to be this way?”,
“could it work any other way?”
Robotics can ask: “what classes of control mechanism are suited for
generating walking behaviors for what leg configurations in what
environments?”
“synthetic approach”
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
2
Robotic Insects
roboinsects: six-legged or eight-legged walking robots have given important
clues to understanding the neural control mechanisms in walking insects
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
3
Robotic Lobster
Joseph Ayers: from biologist to roboticist
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
4
Robotic Fish
MIT “robotuna”
MIT “robopike”
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
5
Robotic models of fish
allow systematic study
of principles governing
efficient swimming.
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
6
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
7
Robotic Brachiation
Robotic models of brachiation used as
test arena to study neural models of
learning of motor control
Other model systems include: snakes,
butterflies, birds, …
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
8
Robotics & Cognitive Science
(why you should care about robots)
Studying real brains:
• since brain produces intelligent
behavior, robotics researches can
learn from better understanding of
brain function.
Biological research inspiring new
architectures/algorithms
Building artificial brains:
• helps us better understand the
problems that the brain is solving
• allows to test cognitive theories
• forces them to be specific
• allows to ask more general questions
When do we “understand” the brain?
When we can rebuild it!?
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
9
Note: the previous argument
highlighted the benefits of
computational modeling.
But why robots?
Embodiment
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
10
Embodiment
Previous argument did not really talk about robots per se, but just the
benefits of computational modeling in general. So why robots?
Brains are not isolated entities but are in continuous interaction with
their environment mediated through their bodies and senses. For a
complete understanding of cognition we need to study this interaction of
brains with their natural, social, and cultural environment.
Therefore, building artificial brains (neural network models) is not enough
but we have to give these artificial brains bodies and environments, too:
“Cognitive Robotics”
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
11
The classic AI approach
perception/
modeling
sensors
planning/
execution
world model
actuators
old AI approach to intelligent behavior
Explain intelligence in terms of symbolic knowledge, logic, search, exact reasoning:
“IF hungry THEN get food”
• discrete symbols for the state of the robot and the world
• planning is analogous to finding a mathematical proof
Problems:
• symbol grounding: how can you extract the symbols from the environment?
• brittleness: systems fail in even slightly different domains
• efficiency: creating and maintaining a comprehensive world model is expensive
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
12
W.G. Walter: Machina Speculatrix
(~1950)
Elsie
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
13
Lesson learned: complex, interesting,
somewhat unpredictable behavior can
emerge when these simple robots interact
with their environment and each other.
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
14
modern implementation with LEGO Mindstorm robots
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
15
V. Braitenberg: “Vehicles” (1984)
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
16
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
17
Braitenberg's primary conclusion in Vehicles is the law
of uphill analysis and downhill synthesis (invention). The
vehicles he describes can produce very complicated
behavior, yet the underlying mechanisms are very
simple. When we try to deduce the mechanisms
controlling these creatures by analyzing their behavior,
we overestimate their complexity. Conversely, we can
invent "downhill," creating creatures that can exhibit
complicated behavior but are composed of only simple
parts.
after Charlie Coglianese, Kushal Dave, and Eric Kennedy
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
18
This vehicle appears to like cold areas and dislike heat.
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
19
light sensor
motorized
wheel
+
+
+
+
Quiz: who shows “fear”, who shows “anger”?
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
20
R. Brooks: “subsumption architecture”
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
21
AI focus: manipulating symbols
new focus: direct interaction with
environment, “subsumption”
“intelligence without reason”
Example: layers of behaviors
locomotion/obstacle avoidance/exploration/map formation/…
“thousands of artificial cockroaches have been built like this”
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
22
Ghengis
Kismet
Problem: this approach does not seem to scale well to designing
human level intelligence.
More modern approaches try to emphasize autonomous learning. The
robot undergoes a developmental period of learning like a child.
Jochen Triesch, UC San Diego, http://cogsci.ucsd.edu/~triesch
23