Research at the Boundary of Robotics
and AI
Prof: Peter Stone
Department of Computer Science
The University of Texas at Austin
AI and Robotics
• Challenge problems
Peter Stone
AI and Robotics
• Challenge problems
• Always on robots (in a human-occupied space)
Peter Stone
AI and Robotics
• Challenge problems
• Always on robots (in a human-occupied space)
• Ad hoc teamwork
Peter Stone
AI and Robotics
• Challenge problems
• Always on robots (in a human-occupied space)
• Ad hoc teamwork
• Our role in a climate where industry is interested
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
How?
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
How?
• Build complete solutions to relevant challenge tasks
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
How?
• Build complete solutions to relevant challenge tasks
• Drives research on component algorithms, theory
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
How?
• Build complete solutions to relevant challenge tasks
• Drives research on component algorithms, theory
• A top-down, empirical approach
Peter Stone
A Goal of AI and Robotics
Robust, fully autonomous
agents in the real world
How?
• Build complete solutions to relevant challenge tasks
• Drives research on component algorithms, theory
• A top-down, empirical approach
“Good problems . . . produce good science” [Cohen, ’04]
Peter Stone
Bottom-Up Metaphors
Russell, ’95
“Theoreticians can produce the AI equivalent of bricks,
beams, and mortar with which AI architects can build the
equivalent of cathedrals.”
Peter Stone
Bottom-Up Metaphors
Russell, ’95
“Theoreticians can produce the AI equivalent of bricks,
beams, and mortar with which AI architects can build the
equivalent of cathedrals.”
Koller, ’01
“In AI . . . we have the tendency to divide a problem into
well-defined pieces, and make progress on each one.
Peter Stone
Bottom-Up Metaphors
Russell, ’95
“Theoreticians can produce the AI equivalent of bricks,
beams, and mortar with which AI architects can build the
equivalent of cathedrals.”
Koller, ’01
“In AI . . . we have the tendency to divide a problem into
well-defined pieces, and make progress on each one.
. . . Part of our solution to the AI problem must involve
building bridges between the pieces.”
Peter Stone
Dividing the Problem
AI
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
The Bricks
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
The Beams and Mortar
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
Towards a Cathedral?
?
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
Or Something Else?
?
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
A Different Problem Division
AI
Peter Stone
Top-Down Approach
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
“Good problems . . . produce good science” [Cohen, ’04]
Peter Stone
Meeting in the Middle
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Peter Stone
Meeting in the Middle
Vision
Game
Theory
Learning
Multiagent
Reasoning
Robotics
Distributed
Optimization
Knowledge
Representation
Natural
Language
Top-down approaches underrepresented
Peter Stone
Choosing the Challenge
• Features of good challenges: [Cohen, ’04]
− Frequent tests; Graduated series of challenges
− Accept poor performance; Complete agents
Peter Stone
Choosing the Challenge
• Features of good challenges: [Cohen, ’04]
− Frequent tests; Graduated series of challenges
− Accept poor performance; Complete agents
• Closed loop + specific goal
• 50-year technical, scientific goals
− Beyond commercial applications — not possible now
− Moore’s law not enough
Peter Stone
Choosing the Challenge
• Features of good challenges: [Cohen, ’04]
− Frequent tests; Graduated series of challenges
− Accept poor performance; Complete agents
• Closed loop + specific goal
• 50-year technical, scientific goals
− Beyond commercial applications — not possible now
− Moore’s law not enough
• There are many — choose one that inspires you
Peter Stone
Good Problems Produce Good Science
Manned flight
Manhattan project
Apollo mission
RoboCup soccer
Goal: By the year 2050, a team of humanoid robots
that can beat the human World Cup champion team.
[Kitano, ’97]
Peter Stone
RoboCup Soccer
• Still in progress
• Many virtues:
−
−
−
−
Incremental challenges, closed loop at each stage
Robot design to multi-robot systems
Relatively easy entry
Inspiring to many
• Visible progress
Peter Stone
AI and Robotics
• Challenge problems
• Always on robots (in a human-occupied space)
• Ad hoc teamwork
• Our role in a climate where industry is interested
Peter Stone
Teamwork
Peter Stone
Teamwork
Peter Stone
Teamwork
• Typical scenario: pre-coordination
− People practice together
− Robots given coordination languages, protocols
− “Locker room agreement” [Stone & Veloso, ’99]
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
• Introduced as AAAI Challenge Problem
[Stone et al. ’10]
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
• Introduced as AAAI Challenge Problem
− Theory: repeated games, bandits
[Stone et al. ’10]
[AIJ, ’11]
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
• Introduced as AAAI Challenge Problem
[Stone et al. ’10]
− Theory: repeated games, bandits
[AIJ, ’11]
− Experiments: pursuit, flocking
[Barrett, Genter, ’12]
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
• Introduced as AAAI Challenge Problem
[Stone et al. ’10]
− Theory: repeated games, bandits
[AIJ, ’11]
− Experiments: pursuit, flocking
[Barrett, Genter, ’12]
− RoboCup experiments;
Peter Stone
Ad Hoc Teams
• Ad hoc team player is an individual
− Unknown teammates (programmed by others)
• Teammates likely sub-optimal: no control
Goal: Create a good team player
• Introduced as AAAI Challenge Problem
[Stone et al. ’10]
− Theory: repeated games, bandits
[AIJ, ’11]
− Experiments: pursuit, flocking
[Barrett, Genter, ’12]
− RoboCup experiments; AAAI Workshops
Peter Stone
AI and Robotics
• Challenge problems
• Always on robots (in a human-occupied space)
• Ad hoc teamwork
• Our role in a climate where industry is interested
Peter Stone