Autonomous Robots & Multi-Agent Systems

Literature Review

M. Schippling
Reverse Engineer
SFI Dynamics of Learning Project

10/29/02
  1. Artificial Intelligence: from Knowledge to Behavior Based
  2. Review of some formative papers
  3. Lego: My History
  4. Machine Learning and Adaptation
  5. Multiple Agents and Collectives
  6. Some Things for the Theory-guys
  7. Useful Publications and Resources

see also: The Dynamics of Learning in Autonomous Robot Collectives website.

 
 

In the beginning there was Artificial Intelligence —
And it was good.

The traditional approach is to build a symantic model of the problem domain in order to deduce "Knowledge". This methodology is useful for Expert Systems; but, robots and other actually useful things just don't work very well.

In the 1980's a number of researchers began developing what is now called:

Behavior Based Artificial Intelligence

Rodney A. Brooks is the progenitor of BBAI.

Taking inspiration from the simple neural control models presented in Valentino Braitenberg's book Vehicles, his reserach at MIT developed a 'Behavior Language' using a subsumption architecture to describe a type of neural connection system (not neural nets) used to program a number of simple robots to perform fairly sophisticated observed behaviors. The abstract of Brooks' paper Elephants Don't Play Chess states the case for BBAI and the paper is an excellent review of his work in the late 1980's.

We have a couple of his other papers online here:


Other players
(idiosyncratic selection process)

Stewart W. Wilson, The Rowland Institute of Science
        The Animat Path to AI
1991 -- Theme setter for the first adaptive learning conference. Proposal for taxonomy and development of "artificial animals" that can survive on their own and build competency from the ground up.
Pattie Maes, MIT
       Behavior Based Artifical Intelligence
1993 -- Short and to the point insider's overview of the field, contrasted with traditional Knowledge Based AI . Good analysis of the strengths and pitfalls of both.
David Payton, Hughes Research Laboratories
        Internalized Plans: A Representation for Action
1991 -- Using plans as resources, not 'infallable' programs. Develops the idea of a flow gradient to direct the robot. The gradients can be modified on the fly depending on circumstances

Ronald C. Arkin, Georgia Institute of Technology
        Integrating Behavioral, Perceptual, and World Knowledge in Reactive Navigation

1991 -- Describes schemas as ways to represent fundamental operations and then generate gradient paths for the robot to follow.

Behavior Based Robotics

1998  -- Excellent review up to mid 1990's.

Luc Steels, Vrei Universiteit Brussel
        Towards a Theory of Emergent Functionality

1991 -- Using distributed, rather than hierarchical, control mechanisms that rely on direct feedback from the environment leads to the emergence of behaviors that are not explicitly programmed. Applies most of the basic dynamical systems theory to Behavior Based Robotics. References Prigogine twice.

Maja J. Mataric, MIT (USC now)
        Navigating with a Rat Brain

1990 -- Biological analogy with the cognitive maps in the rat hippocampus being used to navigate in preference to visual ques. Applied to a real robot, the mechanism uses a topological "map".

Leslie P. Kaelbling, Stanford
        Foundations of Learning in Autonomous Agents

1993 -- Finite state machine theory, including probabilistic state machines. Apparently was not aware of our little Korner of Kaos.

Toys from Industry: Lego

All these guys proposed to use Legos to build lab-bots around 1990:
-> MIT won <-


Machine Learning and Adaptation

The absolute basics:

Ulrich Nehmzow, etal, University of Edinburgh, VUB
        Increasing the Behavorial Repertoire in a Mobile Robot
1992 -- Again, the Edinburgh guys have a good idea: A very simple learning system .

Reinforcement Learning

L.P. Kaelbling, etal, Brown, CMU
        Reinforcement Learning: A Survey
1995 -- Haven't absorbed it.
R.S. Sutton & A.G. Barto
        Reinforcement Learning
1998 -- Standard text on the subject.

Finite State Automata -- TBD

Q-Learning -- TBD

Classifier Systems -- TBD

Neural Nets -- TBD

Genetic Algorithms -- TBD




Multiple Agents and Collectives


Distributed Artifical Intelligence (DAI)

DAI is the 'traditional' name for distributed problem solving. Established in the 1970's, research in the field branched into Multiple Agent Systems (MAS) in the 1990's.

Keith S. Decker, Distributed Problem Solving: A survey (1987)
(I couldn't get it from the durn IEEE because their archive only goes to 1988...)
Divides the arena into four dimensions:

  1. Agent Granularity -- Degree of independence between agents.
  2. Heterogeneity -- How different are individual agents.
  3. Control Mechanism -- Hierarchical vs. distributed, competitive vs. cooperative, etal.
  4. Communication -- None to unlimited, broadcast vs. agent-to-agent, high vs. low level.
Note that our Beowulf array could be situated in the above distributed problem solving matrix as homogenous, more-or-less hierarchical control, with a (currently) somewhat limited inter-agent communication system.

MAS

Nicholas R. Jennings, etal, Queen Mary and Westfield, CMU
        A Roadmap of Agent Research and Development

1998 -- A large review paper that defines the various branches and challenges in the field. They define agent to be distinct from 'helper' programs such as xbiff [paraphrased]:

Gerhard Weiss (ed)
        Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence

2000 -- Haven't read anything but the introduction yet....but, curiously, MIT is not represented in the authors...

The challenges he delineates for DAI generally arise from the micro-macro problem, i.e., the relationship between individual agents and their social environment:


Peter Stone and Manuela Veloso, ATT, CMU
        Multiagent Systems: A Survey from a Machine Learning Perspective

2000 -- A huge, readable, paper with pages of useful references up until the year 2000. They apply the Decker taxonomy to robotic systems along two axes and then describe typical uses and problems for each quadrant of the graph:
  1. Heterogeneity -- From all the same beastie to completely different capabilities and sophistication levels.
  2. Communication -- From only what can be directly observed to broadband transmission.
Their scheme allows the Control Mechanisms to vary from totally independent to the moral equivalent of a single agent with one decision making process (of course, depending on the communication axis).

In any of the quadrants the quality of the agents' interactions can be placed on a scale from cooperative to antagonistic. Their cannonical example of MAS behavior is a Preditor-Prey game which involves agents on both ends of this scale.

The paper points out a number of areas for further research (as of 1999-2000 anyway):

Their not-hidden agenda is to promote Robotic Soccer as a better development arena for MAS.


Pedro Lima, etal, Lisbon Technical University, Georga Tech, etal
        Robo-Cup 2001

2002 -- This is a seemingly frivolous description of the annual IEEE sponsored robot soccer tournament. They even have a division for Eibo Sony-dogs. However the principles behind the play cut to the core of  MAS behaviors and many papers are delivered and software shared amongst competitors (almost all are university engineering/CS classes). [see pages 4 and 9]

They are also starting a Robot Search and Rescue competition. [see page 3]
Just for schadenfreude or something similar, here's the Utah State entry, a modified radio control car....which successfully found victims but failed to communicate their locations and thus was eliminated from competition.


Sample MAS Work

C. Roland Kube & Hong Zhang , University of Alberta
        Collective Robotic Intelligence

1993 -- Box pushing robots cooperate by simply not interfering with each other.

Maja J. Mataric, MIT (now USC)
        Learning to Behave Socially

1994 -- Learning to cooperate by "vicariously" sharing behavior rewards.

Cao, Y. Uny, etal, UCLA, Caltech
        Cooperative Mobile Robotics: Antecedents and Directions

2000 -- Yet Another Massive Review paper....more references...specifically Asian authors?

Robert Grabowski, etal, CMU
        Heterogeneous Teams of Modular Robots for Mapping and Exploration

2000 -- Describes small home-brew robots with differing capabilities that work in groups ( Millibots , developed at CMU). Need to study this one.
Claude F. Touzet, Oak Ridge
        Robot Awareness in Cooperative Mobile Robot Learning
2000 -- Using a hierarchy of knowledge of neighbor robots to limit state space. The further away the neighbors are, the less you need to remember about them.
Dieter Fox, etal, CMU, etal
        A Probabilistic Approach to Collaborative Multi-Robot Localization
2000 -- Using multiple robots to increase their localization accuracy by sharing relevant information.

Robin R. Murphy, etal , University of South Florida
        Emotion-Bassed Control of Cooperating Heterogeneous Mobile Robots

2002 -- Describes a two robot system where each perform different tasks that require a rendevous. The 'emotional' content is how confident they are that the rendevous will be on schedule. Given their emotional states, the 'bots can modify their behavior to accomplish their tasks.

S.S. Ge & Y.J. Cui, National University of Singapore
        Dynamic Motion Planning for Mobile Robots Using Potential Field Method

2002 -- Tracking a moving target with moving obstacles, simulated and real 'bots. [see page 14]

Some Things for the Theory-guys:


Useful Publications and Resources

Books and Collections

Vehicles: Experiments in Synthetic Psychology, 1984; Braitenberg, V. -- Fun thought experiments you can now do with the Lego Mindstorms kit.

Designing Autonomous Systems, 1990; Maes, P. (ed) -- Reprints from journal Robotics and Autonomous Systems, vol 6.

From Animals to Animats #1,2,3, 1991,3,4 -- Proceedings of the first three International Conferences on Simulation of Adaptive Behavior, many interesting papers in each issue.

Toward Learning Robots , 1991 -- Reprints from journal Robotics and Autonomous Systems, vol 8.

Reinforcement Learning , 1998; R.S. Sutton & A.G. Barto -- Standard text.

Multiagent Systems , 2000; G. Weiss (ed) -- Collected papers on MAS.

Proceedings of the First International Conference on Multi-Agent Systems, 1995, AAAI -- Formative work.

Adaptation, Coevolution, and Learning in Multi-Agent Systems, 1996; AAAI spring symposium.

Adaptation and Learning in Multi-Agent Systems, 1996; Weiss and Sen -- TBA.

An Introduction to AI Robotics, 2000; Murphy, R. -- On order.

Robot Teams: From Diversity to Polymorphism, 2002; Balach and Parker -- Ordered

Journals

IEEE {Journal of, Transactions on} Robotics and Automation --
http://ieeexplore.ieee.org/Xplore/DynWel.jsp

Autonomous Robots --
http://www.kluweronline.com/issn/0929-5593/current

Machine Learning --
http://www.kluweronline.com/issn/0885-6125/current

Autonomous Agents and Multi-Agent Systems --
http://www.kluweronline.com/issn/1387-2532/current

Robotics and Autonomous Systems --
http://www.sciencedirect.com/science

And, of course...

Our own Dynamics of Learning website: