To be updated !!

TOYBOTS

Towards an Artificial Pet

A typical LEGO Toybot.

Newspaper article about the Toybot project (at Robotix'97) from The Guardian: Trouble in Toytown, The Guardian, 20 March 1997
and one from the Swiss magazine Facts: Künstliche Haustiere -- Fifi mit Festplatte --> -- wedelt auf Kommando. Facts. 21 May 1997.
Research Staff:

ALife and Adaptive Robots Group, The Danish National Centre for IT Research -- CIT:

Department of Artificial Intelligence, University of Edinburgh, UK: Institute of Psychology, National Research Council, Rome, Italy:

A strong limitation of traditional robot kits is the assumption that to build an intelligent toy-robot (toybot), it is better to program the robot as a sort of ``computer with wheels''. Our project is based on a different point of view: we consider a mobile toy robot as a little pet.

Children (and, in general, human beings) play with animals by interacting with the pet's behaviour. For example, different forms of interaction between children and pets are:

  • putting a pet in a new environment
  • putting a pet in ``competition'' or ``cooperation'' with other pets
  • training a pet to perform a task

    In all of these forms of interaction, children are not interested in ``programming'' the pet brain. They are not interested in ``how the brain (or the mind) works''. In short, human beings consider animals as autonomous systems with a specific and individual way to behave. Therefore, in our approach, the children-toybots interaction focuses only on the ``behaviour''. In order to permit children to build interesting mobile robots, we try to avoid any form of ``information processing methodology'' (programming).

    Artificial Life techniques can represent useful solutions for our goals. For a specific task designed by the child, we allow an evolutionary process to design the appropriate robot body plan, that the child can assemble. In co-evolution and/or through a user-guided genetic algorithm, the controller of the robot is evolved and down-loaded to the robot. The user-guided genetic algorithm consists of representations of the robots' behaviours on the screen, and the user performs the selection based on the user's own judgement. Re-training in the real world can be done through reinforcement learning.

    Screen shot of the simulator to that uses a user-guided genetic algorithm to evolve LEGO robots.

    The toybots' behaviours are controlled by Neural Networks that model the brain of pets. One main advantage of using Neural Networks is that they can be trained to solve specific tasks -- in the case of toybots, the Neural Networks can be trained to learn specific behaviours.

    At the current state of the project, children can develop simple behaviours such as obstacle avoidance, following or avoiding lines on the floor, attraction or fear of light, etc. for LEGO robots. The idea is to develop this futher, so that children can teach pet-robots different skills and have them to interact with each other.

    The TOYBOTS project is a research collaboration between the Mobile Robots Group at Department of Artificial Intelligence, University of Edinburgh, UK and the Institute of Psychology at National Research Council, Rome, Italy.

    Other photos (click to enlarge):

    Contact:

    Henrik Hautop Lund
    The Danish National Centre for IT Research -- CIT
    University of Aarhus
    Ny Munkegade, bldg. 540
    8000 Aarhus C.
    Denmark
    e-mail: hhl@daimi.aau.dk.

    Henrik Hautop Lund

    Mobile Robot Group at University of Edinburgh