Club of Amsterdam, Future, Think Tank ..
Menu

Club of Amsterdam Journal
Free Subscription
click here





































 

 

 


 

 

 

 


 


 

 


 

 



 


 

 


 

 

 

 



 




 






 


 

 

 

 


 

 

 


 

 

 

 


 

 

 


 

 


 

 

 

 

 

 

 


 

 

 

 

 

 


 

 



 

 

 

 

 

 


 

 

 



 

 

 

 

 

 


 

 

 



 

 

 

 

 

 


 

 

 

 

 

 


 

 

 

 

 


 

 

 

 


 

 

 

 

 

 

 

 


 

 

 

 


 

 

 

 

 

 

 

 


 

 

 

 

 

 

 

 





keyword search

Emerging shared action categories in robotic agents through imitation
Average reader rating: 0  
by Jansen, De Vylder, de Boer, Belpaeme 21 the future of Robotics

1 Introduction
Most of the work that has been published on imitation in robots focuses on the learning of action categories in a teacher - student context (Vogt, 2000; Billard and Hayes, 1997; Alissandrakis et al., 2002) . In such a set-up one agent acting as teacher already has action categories. By observing the teacher who is executing actions, the action categories can be passed on to the student: by imitating the teacherís action using for instance inverse kinematics and evaluating that action, the learner can know whether he correctly reconstructed the observed action. However, such a set-up does not explain how action categories emerge. How does the teacher acquire its categories when they are not preprogrammed by a human operator?

We propose a set-up in which new action categories can emerge when imitation of actions fails. This is done in a population of agents engaging in imitative interactions, called imitation games. Action categories are only learned if they can be successfully imitated. If an action is hard to observe or to imitate, it will not be learned by other agents. The experiments are conducted both in simulation and on real robots, however results presented in this article were only obtained in simulation. Our concept of imitation games strongly resembles the concept of imitation games used in (de Boer, 2000) in the context of vowel systems. The imitation game presented in this paper is work in progress.

In section two, our experimental set-up is proposed. Section three presents the actual imitation game, section four proposes objective measures for determining how successful the imitation game is and results are presented in section six. Future work is discussed in section seven.

[...]

8 Conclusion
In this paper we propose an experimental set-up, based on imitation, suited for conducting experiments on the exact conditions required for shared action categories to emerge in a population of real-world agents. We have shown that these categories can be shared in the population without a fixed teacher - student pattern of imitation where the teacher already has fully developed categories. Instead, the action categories can emerge and become shared through multiple imitation games in a population of agents where all agents can act as a teacher or student. The learnt actions can be observed, categorized and imitated by other agents, which is not guaranteed in set-ups where a teacher starts with built-in action categories and transfers those categories to the student(s).

Preliminary simulation results show that using imitation games, shared repertoires of action categories can be obtained. These repertoires are non-trivial, as they consist of multiple action categories. They are also very successful: for populations of only two agents imitation success is always in the 90% range even though noise is present. In larger populations the imitation success is lower, but still better than random repertoires would achieve. Imitation games in this setup are based on the imitator of the imitation game updating its action categories. In this case calibration, inverse kinematics and non-verbal feedback are required.

If the population is restricted to only two agents, action categories can emerge and be shared without the action space and the observation space are calibrated, without feedback between the agents about the outcome of the game and without the agents have built-in notion of the inverse kinematics of their manipulator.

Download the full article [*.pdf]: Click here


Visit also the conference about 'the future of Robotics' and the sections with books, articles and links.









Rated 0 by other users. What do you think? [rate this article]

Copyright © 2002-2020 Club of Amsterdam. All rights reserved.    Contact     Privacy statement    Cancellation Policy