Fundamental Research in Machine Learning for Robotics
The Artificial Intelligence group (AI Lab) is the fundamental
research unit of SoftBank Robotics Europe. We focus on developmental
robotics, with the goal to make progress in the understanding and
modeling of the mechanisms of development and learning in robots.
Learning from Demonstration
Enabling users to teach their robots new tasks at home is a major
challenge for research in personal robotics. Dynamic Movement
Primitives provide a promising basis for building a learning system
that could be employed for such a task.
These two videos show how Pepper can be taught new skills in
relatively short time. We use a "learning by demonstration" procedure,
where a movement is first demonstrated to the robot by guiding its
arm.
From there, Pepper has to improve its performance through
trial-and-error learning. Even when the initial demonstration is not a
successful task execution, Pepper can still learn the skill.
The movement is optimized using simple evolutionary search (CMA-ES). Our
implementation uses the freely available software library dmpbbo.
Gaining Control over the own Body through Trial-and-Error
Learning
In this work done by my PhD student
Pontus Loviken,
the robot Nao learns to take control of its own body. It is given no
information about the structure of its body, and has to figure
everything out strictly through trial-and-error learning. Its goal is
to be learn how to roll its body (one of the very first locomotion
behaviors infants acquire in their development).
In contrast to standard Reinforcement Learning techniques to learn this
type of robot behavior, the learning system in this case efficiently
accumulates knowledge (in the form of a model) about the way its own
body behaves in the environment. This is beneficial, as it allows the
robot to exploit the gained knowledge to produce a multitude of
behaviors, whereas Reinforcement Learning only optimizes a single
behavior.
Anna-Lisa Vollmer, andNikolas J. Hemion.
"A User Study on Robot Skill Learning Without a Cost Function:
Optimization of Dynamic Movement Primitives via Naive User
Feedback."
Frontiers in Robotics and AI, vol 5, article 77.
2018.
Pontus Loviken,Nikolas J. Hemion,Alban Laflaquière,Michael Spranger, andAngelo Cangelosi.
"Online Learning of Body Orientation Control on a Humanoid Robot
using Finite Element Goal Babbling."
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS).
2018.
Pontus Loviken, andNikolas J. Hemion.
"Online-Learning and Planning in High Dimensions with Finite
Element Goal Babbling."
Joint IEEE International Conference on Development and Learning and
Epigenetic Robotics (ICDL-EpiRob).
2017.
Nikolas J. Hemion.
"Discovering Latent States for Model Learning: Applying
Sensorimotor Contingencies Theory and Predictive Processing to
Model Context."
ArXiv paper.2016.
Nikolas J. Hemion.
"Context Discovery for Model Learning in Partially Observable
Environments."
Joint IEEE International Conference on Development and Learning and
Epigenetic Robotics (ICDL-EpiRob).
2016.
Thomas R. Colin,Tony Belpaeme,Angelo Cangelosi, andNikolas J. Hemion.
"Hierarchical Reinforcement Learning as Creative Problem Solving."
Robotics and Autonomous Systems, vol 86, pp 196–206.
2016.
Alban Laflaquière, andNikolas J. Hemion.
"Grounding object perception in a naive agent's sensorimotor
experience."
Joint IEEE International Conference on Development and Learning and
Epigenetic Robotics (ICDL-EpiRob).
2015.
Alban Laflaquière,Nikolas J. Hemion,Michaël Garcia Ortiz,and Jean-Christophe Baillie.
"Grounding Perception: A Developmental Approach to Sensorimotor
Contingencies"
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS).
2015.
Nikolas J. Hemion.
"On learning and representations in cognitive architectures for
HRI"
Workshop on Cognitive Architectures for Human-Robot Interaction,
ACM/IEEE International Conference on Human-Robot Interaction.
2014.
Nikolas J. Hemion,
"Building Blocks for Cognitive Robots: Embodied Simulation and
Schemata in a Cognitive Architecture."
Ph.D. dissertation, Technische Fakultät, Universität Bielefeld.
2013.
Nikolas J. Hemion,Frank Joublin,and Katharina J. Rohlfing.
"Integration of sensorimotor mappings by making use of
redundancies."
IEEE International Joint Conference on Neural Networks (IJCNN).
2012.
Nikolas J. Hemion,Frank Joublin,and Katharina J. Rohlfing.
"A competitive mechanism for self-organized learning of
sensorimotor mappings."
IEEE International Conference on Development and Learning (ICDL).
2011.