Transcript What is a robot? - Duke Computer Science

Artificial Intelligence
•
•
•
Can computers/machines think?
 Many (most?) AI researchers don’t care
What is AI then?
 Systems that think like humans
 Systems that act like humans
 Systems that think rationally
 Systems that act rationally
 Any more?
An very active interdisciplinary any area of research encompassing robotics,
machine vision, machine learning, planning, reasoning under uncertainty,
natural language processing, knowledge discovery and representation, and
many more
CPS 001
AI.1
Applications of Artificial Intelligence
•
•
•
•
•
Robotics
 Nursebot , Honda ASIMO, Roomba
Decision support
 Microsoft’s Lumiere (1998), Collaborative Filtering
Game Playing
 Deep Blue (1997), TD-Gammon (1995)
Machine vision/graphics
 Campanile, Medical research(Fluoroscein)
Lots of practical work in AI
 100 patents in AI in 1989
 2200 classified as AI in 1999 (5600 mention AI)
CPS 001
AI.2
Control basics
•
•
•
Open-loop control
 Compute trajectory a priori and make necessary actions to
complete task
Closed-loop control
 Use sensors to provide feedback to modify the trajectory
and actions
Degrees of freedom
 Number of dimensions to precisely represent an object
 Most object have 6 degrees of freedom
• (x,y,z) translational coordinates
• (f,q,g) Euler/rotational coords - (pitch,yaw,roll)
CPS 001
AI.3
Agents and Environments
CPS 001
AI.4
Autonomous Vehicle Control:
The World
•
•
Noisy, unpredictable and hostile
CPS 001
Delayed
feedback from actions
•

Partially Observable
Significant challenge for AI
AI.5
Reflex agent
CPS 001
AI.6
Reflex agent with state
CPS 001
AI.7
Goal-oriented agent
CPS 001
AI.8
Utility-based agent
CPS 001
AI.9
Autonomous Vehicle Control:
Approaches
•
•
Vehicle control
 Significant progress on low-level sensing and control
 [Dickmans and Zapp, 1987; Godbole and Lygeros, 1993;
Pomerleau, 1993; Malik et al., 1997]
High-level reasoning
 SAPIENT system at CMU
 RL methods
 The Bayesian Automated Taxi Project
• Goal: Develop and test control architecture for learning control
of autonomous vehicles
• Domain: Simulator
– Much safer!
CPS 001
AI.10
Autonomous Vehicle Control:
The BAT Project
•
The Bayesian Automated Taxi (BAT) Project architecture
Modeling Traffic
Probability
Dynamic Bayesian Networks
Sensor Data
State.t-1
State.t
Inference
Sense.t
xi
State
Lateral &
Longitudinal
Control
a
CPS 001
&
Features
Decision Making
Action selection
Actions
High-level
Control Parameters
AI.11
Autonomous Vehicle Control:
Learning to Drive

No single optimal trajectory or path
 Not easily amenable to supervised learning or regulatory control methods
Developed an explicit policy representation for control which performed
robustly in a number of driving scenarios
 Somewhat fragile and not easily adaptable
•
Reinforcement learning (RL)
•

•
Successively improves and adapts control strategies through trial-and-error
interactions with a dynamic environment
RL techniques have shown some promise in solving complex control
problems
 TD-Gammon [Tesauro, 1992], Cell phone routing [Singh, 1996]
 Need to scale and extend for continuous control domains
CPS 001
AI.12
What is a robot?
•
Definitions
 Webster: a machine that looks like a human being and
performs various acts (as walking and talking) of a human
being
 Robotics Institute of America: a robot is a reprogrammable
multifunctional manipulator designed to move material,
parts, tools, or specialized devices through variable
programmed motions for the performance of a variety of
tasks
 What’s our definition
•
Components of a robot system?
CPS 001
AI.13
History
•
•
•
•
•
1921: Karel Capek’s play, Rossum’s Universal Robots
1942: Asimov wrote Runaround which contained the “Three
Laws of Robotics”
1. A robot must may not injure a human being or through
inaction allow a human being to come to harm.
2. A robot must obey the orders given it by human beings,
except where such orders would conflict with Rule 1.
3. A robot must protect its own existence, as long as such
protection does not conflict with Rules 1-2.
1948: Weiner wrote “Cybernetics”
1961: General Motors’ puts UNIMATE online (first industrial
robot)
1970: SRI’s Shakey: first AI mobile robot
CPS 001
AI.14
Uses of robots
•
Where and when to use robots?
 Tasks that are dirty, dull, or dangerous
 Where there is significant academic and industrial interest
Ethical and liability issues
•
What industries?
•
What applications?
•
CPS 001
AI.15
Robots Good or Bad
QuickTime™ and a YUV4 20 code c d eco mpres sor a re ne eded to see this picture .
CPS 001
AI.16
The Nursebot Initiative
Quic kTime™ and a YUV420 codec decompress or are needed to s ee this picture.
Quic kTime™ and a YUV420 codec decompress or are needed to s ee this picture.
CPS 001
Sebastian Thrun with Greg Armstrong, Greg Baltus,
Jacqueline Dunbar-Jacob, Jennifer Goetz, Sara Kiesler,
Judith Matthews, Colleen McCarthy, Michael
Montemerlo, Joelle Pineau, Martha Pollack, NicholasAI.17
Roy, Jamie Schulte
Teleoperation
•
•
•
•
•
Remote control of robot manipulators
Intermediate solution to tasks that currently could or should not be
handled with autonomous robots
Various levels of abstraction
Issues
 Control latency
 Communication bandwidth
 Operator training
Some definitions:
 Control system: arrangement of physics components connected or
related in such a manner as to form and/or act as an entire unit
 Kinematics: the description or study of the geometry of motion
 Dynamics: the description or study of the forces that affect the
motion of objects
CPS 001
AI.18
Telepresence
•
Be there without really being there
Simple examples?
•
PRoP
•
CPS 001
AI.19