[robotics-worldwide] PhD Openings in Machine Learning, Robot Control and Human-Robot Interaction at the Department of Advanced Robotics, Italian Institute of Technology

Sylvain Calinon Sylvain.Calinon at iit.it
Fri Sep 3 06:19:58 PDT 2010


The Department of Advanced Robotics (http://www.iit.it/en/advanced-robotics) at the Italian Institute of Technology IIT (an English language Institute) located in Genoa has a number of PhD openings within the research areas of Machine Learning, Robot Control and Human-Robot Interaction (starting in January 2011). Please see below the list of the available themes.

--------
Theme 2.7: Incremental learning for compliant robots through the superposition of basis force fields

The aim of this PhD proposal is to study incremental learning techniques applied to compliant robots. The advantage of moving from offline to online learning is twofold. Compared to stiff robots, compliant robots allow users to get close to the robot, sharing not only the same environment but also the same tasks. The development of compliant robots facilitates the demonstration of new skills by using kinesthetic teaching (e.g., by using gravity compensation to move the robot as if it had no weight). When working in collaboration with robots, this type of interaction increase the quantity of data collected by the robots, and it becomes important to consider demonstration and reproduction as an interlaced process where the robot gradually refines its skill.

Models based on a superposition of basis force fields have interesting characteristics for online learning, since most of the computation can be performed in parallel (each basis force field can be refined independently). Exploiting the parallel computation power of Graphics processing unit (GPU) for Machine Learning purpose is a recent but rapidly growing trend of research that could provide robots with cheap and portable parallel processing capabilities. Another route is to consider learning as both a continuous and disruptive adaptation process. Incremental learning techniques consist most of the time of progressive and continuous adjustments that allow the system to handle perturbations and slowly varying environments and adaptation to the robot's capabilities. However, as their biological counterparts, such systems would also occasionally benefit from disruptive adaptation mechanisms. Cubs can optimize their crawling skills until their legs are strong enough to explore a completely different way of moving on two legs. It is proposed to study robot learning strategies allowing to acquire and refine skills through such a combination of imitation, exploration, evaluation of performance and practice.

For further details, please contact: sylvain.calinon at iit.it

--------
Theme 2.8: From trajectory learning to force skills acquisition in compliant robotic arms

Kinesthetic teaching offers a user-friendly and intuitive method to teach robots new skills. Compliant robots offer new perspectives in this direction, where torque feedback can be used to let the user move the robot as if it had no weight, and as if no motors were in its articulations. Through gravity compensation actuation, the user can concentrate on the task to demonstrate by grasping and re-grasping the robot in ways that are natural and efficient for him/her to execute the task. This proposal aims at extending the range of skills that a robot can learn by moving the representation of the task constraints to the force domain. It is first proposed to explore how robot learning capabilities can be improved through the use of both torque information at the level of the joints and force sensing at the level of the end-effector. It is then suggested to explore the simultaneous consideration of multiple constraints in the force domain (e.g. by considering multiple contact points), which would offer new perspectives in physical human-robot interaction.

For further details, please contact: sylvain.calinon at iit.it

--------
Theme 2.9: Safe human-robot interaction

Traditional and new markets ranging from the industrial field to the service robotics, both for professional and domestic applications show an increasing desire for innovative robots able to collaborate with humans. One of the current challenges in robotics is therefore to improve safety in human-robot interaction. It is well known that collaborative tasks need to be performed in a flexible manner, where humans work side-by side with robots sharing workspace and objects. The risk assessment used for industrial applications guarantees safety around the robot by drawing a no go area where motion of the robot is halted if a human intrusion is detected. This approach, however, is far too restrictive for tasks where humans are required to closely interact with robots. As a consequence, considerable research is conducted, both at the hardware and software level to decrease the risk while improving the safety in human-robot interaction. At the hardware level, compliant actuators have been proposed as a safer approach for driving robots which have to interact with humans. Using compliant actuators it is possible to reduce the forces of an impact between a robot and a human, in the cases of an unexpected collision. Within the long-term project to develop a compliant humanoid robot (AMARSI, European project), one open PhD position is available. The aim of this project is to develop a control strategy for a multi-dof compliant robotic system along with the analysis of the optimal stiffness and the input trajectory profiles for minimizing energy consumption. At the software level, control strategies for adjusting the robot's behavior on the basis of the detected level of hazard in human robot interaction are required. The aim of the project is to first define a risk indicator on the basis of the safety restrictions (user or robot side) depending on the task to accomplish. This is necessary to continuously let the robot be aware of the environmental changes with respect to its internal state. The project will investigate how to construct the risk indicator and which are the parameters to take into account for describing changing safety constraints for several different tasks. Task redundancy and user intention will mainly be considered for this purpose. Then a control strategy will be implemented to accomplish the required tasks on the basis of the current level of risk. The project includes the use of the KUKA LWR and/or Barrett WAM robotic arm.

The candidates should have a background in Engineering, Mathematics or Physical Sciences (control theory, system modeling) and computer science. Required technical skills:  Matlab, C/C++.

For further details, please contact: irene.sardellitti at iit.it

--------
Application Requirements:

Interested candidates holding a Master degree in Mechatronics, Computer Science, or other related fields are invited to apply for admission. Applicants should ideally have strong competencies in one or more of the followings areas: machine learning, robot control, MATLAB and C/C++ programming. International applications are encouraged and will receive logistic support with visa issues, relocation, etc.

To apply please send a detailed CV, a statement of motivation, and any additional support material (reference letters, degree certificates, grade of transcripts) to Sylvain Calinon (sylvain.calinon at iit.it), Irene Sardellitti (irene.sardellitti at iit.it) and Floriana Sardi (floriana.sardi at iit.it). In addition, the applicants should fill the online application form at http://www.iit.it/en/home.html under the "PhD School - Life and Human Technologies" Tab no later than the 24th of September 2010.





More information about the robotics-worldwide mailing list