Skip to content. | Skip to navigation

Personal tools
Document Actions

Project Description

Posters and docs presenting the project goals and topics at various levels of detail, both for the wide public and for the scientist

Key documents and presentations summarising the project goals and results

IMPORTANT NOTICE: The material reported in this web-page is copyrighted; you can use it integrally or in parts to the extent that you cite the source!

A1 poster: A colourful scketch poster on the project vision, core working hypothesis, partners, and background ideas of the project presented at EpiRob09
14/11/2009, Venice (Italy), EpiRob2009
IM-CLeVeR - Instrinsically Motivated Cumulative Learning Versatile Robots
Download .pdf file

2-pages extended abstract: An abstract on the project goals published in the EpiRob09 proceedings (the contents below, after 'summary', are based on this document).
14/11/2009, Venice (Italy), In Canamero L., Ouderyer P.-Y., Balkenius C. (eds.). Proceedings of the Ninth International Conference on Epigenetic Robotics, pages 189-190, Lund University Cognitive Studies n. 146, Lund: Lund University.
Baldassarre G. et al. (2009), The IM-CLeVeR Project - Intrinsically Motivated Cumulative Learning Versatile Robots.
Download .pdf file

35-page presentation: An invited presentation (EpiRob09) on the project vision, plus examples of main previous work done by partners on the project topics
14/11/2009, Venice (Italy), EpiRob2009
Gianluca Baldassarre, IM-CLeVeR - Intrinsically Motivated Cumulative Learning Versatile Robots.
Download .ppt file     Download .pdf file (limitation: many overlapped pictures and text due to missing animations)

20-page presentation: An invited presentation on the project vision, with some examples of the work carried out at CNR-ISTC
18/01/2010, Sestri Levante (Italy), The iCub Workshop,
Gianluca Baldassarre, The IM-CLeVeR Project - Intrinsically Motivated Cumulative Learning Versatile Robots.
Download .ppt file     Download .pdf file (limitation: many overlapped pictures and text due to missing animations)

43-page presentation: A presentation of the Coordinator on the project progress from 05-2010 to 04-2011.
01/07/2011, Lugano (Switzerland), Second Review Meeting.
Gianluca Baldassarre, The IM-CLeVeR Project - Progress of last reporting year.
Download .pdf file

5-page document: A summary by the Coordinator of the IM-CLeVeR Project overall formal issues, objectives, resources, partners, and main goals and achievements, dated 04 July 2012
Download .pdf file

43-slide presentation: A presentation by the Coordinator on the main achievements of the IM-CLeVeR Project in the first two years of teh project (up to April 2011).
Download .pdf file

21-slide presentation: A presentation by the Coordinator on the main achievements of the IM-CLeVeR Project in the first three years of the project (up to July 2012).
Download .pdf file

7-page document: List of seminars were held at the Institue of Cognitive Science and Technlogies, National Resaerch Council, Rome on the topics of the project, during October-December 2013. For each seminar the document reports: name of speaker(s), title and abstract of the presentation, a key reference when available.
Download .pdf file

Summary

In the rest of this web page you will find a rather detailed presentation of the core ideas of the IM-CLeVeR Project, based on the extended abstract presented at EpiRob2009 (see above).

IM-CLeVeR aims at developing a new methodology for designing robot controllers that can: (a) cumulatively learn new skills through autonomous development based on intrinsic motivations, and (b) reuse such skills for accomplishing multiple, complex, and externally-assigned tasks.

This goal will be pursued by investigating three fundamental issues:

  • The mechanisms of abstraction of sensorimotor information;
  • The mechanisms underlying intrinsic motivations;
  • Hierarchical architectures that permit cumulative learning.

The study of these issues will be conducted on the basis of empirical experiments run with monkeys, children, and human adults, with bio-mimetic models aimed at reproducing and interpreting the results of such experiments, and through the design of innovative machine learning systems.
The  models, architectures, and algorithms so developed will be validated with experiments and demonstrators run with the simulated and real iCub humanoid robot.

Introduction

How can we create truly intelligent and autonomous machines and robots? This goal has both a huge technological and scientific importance.

As a technology, autonomous intelligent machines can be exploited, for example, to perform repetitive tasks that humans do not like to carry out and conduct missions in hostile environments.

On the scientific side, the ability to construct truly intelligent machines can shed new light on the mechanisms underlying learning and intelligence of humans and other primates, thus also enabling better treatments of psychiatric and neurological disorders.

The IM-CLeVeR project aims at developing a new design methodology for building autonomous intelligent robots based on intrinsically-motivated cumulative learning of skills.

The central idea behind this new design methodology is that, instead of directly programming, training or evolving a set of specific skills in robots, we should endow them with developmental programs that allow an autonomous development of the needed skills on the basis of prolonged periods of interactions with the environment under the guidance of intrinsic motivations. Robots could then use the general abilities so acquired as building blocks for the solution of tasks that are relevant for the robot's users.

Notice how these types of processes mark some of the most intelligent aspects of complex organisms' behaviour, in particular human and non-human primates. For example, children at play carry out several activities driven only by intrinsic motivations such as curiosity. These activities allow them to acquire knowledge and skills exploited in later adult stages to pursue useful goals. The main objectives of the project will be pursued with these phenomena in mind.

The Central Working Hypothesis of the IM-CLeVeR

The central working hypothesis of the project is that cumulative, open-ended learning in artificial systems must be based on three fundamental principles:

  • Hierarchical architectures. Cumulative learning architectures for controlling robots should have the capability of developing sensorimotor and cognitive skills in an incremental hierarchical fashion. This requires: (a) acquiring skills and systematically increasing their complexity; (b) learning new skills using previously acquired skills as building blocks; (c) storing new skills without forgetting (and possibly improving) previously acquired ones.
  • Novelty detection and intrinsic motivations. A cumulative learning robot needs internal drives that focus learning on skills that: (a) are novel for the robot; (b) are within the robot's `zone of proximal development' -- that is the robot has the drive to learn new skills that can be acquired on the basis of those already in its repertoire. To achieve this, the robot should be endowed with `intrinsic motivations' that lead it to autonomously engage in activities that produce the maximum learning rate and/or information gain. Internal motivations differ from external motivations and rewards as the latter are associated with the practical outcomes that actions produce on the external world (e.g., food or sex in organisms or accomplishment of users' goals in robots). Intrinsically motivated learning must rely on `novelty detectors', devices capable of monitoring and measuring the level of subjective novelty of action outcomes and learning rates so as to focus robot's activity on suitable experiences and boost learning speed.
  • Sensory abstraction and attention. Although sensory abstraction is a widely-investigated topic in cognitive sciences (e.g. in computer vision), the project will aim at isolating and studying the particular problems of abstraction related to the specific topics of the project, namely novelty detection and hierarchical architectures for cumulative learning.

 

Objectives of IM-CLeVeR

IM-CLeVeR has four main scientific and technological objectives:

  • To advance our knowledge about how cumulative learning is achieved in natural organisms. To this purpose, the project involves the implementation of empirical non-invasive experiments on intrinsically motivated learning in monkeys, children, human adults, and Parkinson patients, on the basis of novel experimental paradigms suitable for studying exploration, novelty detection, and the (cumulative) acquisition of novel actions".
  • To advance our knowledge about the mechanisms underlying intrinsically motivated cumulative learning in natural organisms. To this purpose, the project will develop bio-mimetic models (including both computer simulations and robotic experiments) aiming at reproducing and explaining the empirical findings provided by the aforementioned empirical experiments. In addition to its scientific value, this effort will also allow isolating new computational principles exploitable in robots.
  • To develop new machine learning techniques, architectures, and learning algorithms for the optimal design of cumulative learning robots. In particular, the project will aim at making substantial progress in the three distinct but related principles of the project working hypothesis: (a) hierarchical architectures, (b) intrinsic motivations, and (c) perceptual abstraction and attention.
  • To integrate the knowledge gained by the empirical experiments, the bio-mimetic computational models developed to interpret them, and the machine learning architectures and algorithms for building real robots demonstrating cumulative learning abilities. This challenge will involve the use of three iCub humanoid robotic platforms for the development of two demonstrators: CLEVER-K, a technologically-oriented demonstrator that will be tested in a kitchen scenario, and CLEVER-B, a demonstrator with which will be used to reproduce and interpret the results of the experiments carried out with monkeys and children.