Internships and Master Theses
If you are a student at Universidad de Zaragoza, the Academia General Militar or you want to do an internship at the Spanish University Center of Defense, you might be interested in one of these or other projects. For any question or inquiry, contact me through email:
rmcantin (at) unizar.es
Please note that, at this moment, I cannot provide any financial support for internships. If you are interested in getting a PhD or working as a Postdoc, send me you CV and it will be automatically included in the selection as soon as a new position is available.
Also, while contacting me, tell me which kind of project would you be interested in. It could be one of those listed here or your own idea.
Robot grasping using active learning
Grasping novel objects is a process that is made in two steps: First a visual inspection give us a set of possible grasping surfaces. Then, we optimize the motor primitives of our hands to achieve the most stable grasp. We can even try different grasps, accommodating the hand in different postures until we find the optimal one [1].
Recent advances in active learning for robotics allow to speed-up the search process using Bayesian reasoning. The visual information give a prior information over the optimal grasping, while each trial can be used to update the knowledge about the location of the optimal grasp. The student will be using an existing library for the search process [2].
[1] Kroemer, O., R. Detry, J. Piater and J. Peters: Active Learning Using Mean Shift Optimization for Robot Grasping. Proceedings of the 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2009)
[2] Ruben Martinez-Cantin, Nando de Freitas, Eric Brochu, Jose Castellanos and Arnaud Doucet (2009) A Bayesian Exploration-Exploitation Approach for Optimal Online Sensing and Planning with a Visually Guided Mobile Robot. Autonomous Robots - Special Issue on Robot Learning, Part B, 27(3):93-103.
The student should have knowledge of C/C++. Experience in Linux is recommendable.
Imitation learning for robots and virtual actors
The problem of designing complex motion primitives is analogous in humanoid robotics and virtual actors. Natural motions are typically extracted from some kind of motion capture system using a real actor or teacher. Then, the robot or the virtual actor has to adapt the data to his own kinematic structure, imitating the behavior of the teacher.
The student will test different algorithms [1,2] for learning from modeling human motion. The algorithms will be evaluated in simulation and using a humanoid robot.
[1] Gaussian Process Dynamical Models for Human Motion J. M. Wang, D. J. Fleet, A. Hertzmann. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI). Feb 2008. Vol. 30, No. 2. pp. 283-298.
[2] Modeling Human Motion Using Binary Latent Variables Graham Taylor, Geoffrey Hinton, and Sam Roweis. Proc. of Advances in Neural Information Processing Systems (NIPS) 19, 2007
The student should have knowledge of Matlab and/or C/C++.
Actively learning affordances
Affordances encode relationships between actions, objects, and effects. They play an important role on basic cognitive capabilities such as prediction and planning, which are the roots of developmental learning. They can also be used to generalize object recognition usign not only visual information.
In order to learn the affordances, the robot has to interact with the objects and check the effects of every action. This is a time consuming process that requires interacting with different objects in different ways. In this project, the student will develop methods to find the most interesting actions to be executed in order to speed up the learning process.
[1] Learning Object Affordances: From Sensory Motor Maps to Imitation, Luis Montesano, Manuel Lopes, Alexandre Bernardino, José Santos-Victor, IEEE Transactions on Robotics and Automation, 2008.
The student should have knowledge of C/C++. Experience in Linux is recommendable.
Global optimization for robot learning
Bayesian optimization is a powerful technique that has recently attracted the attention of the robotics and artificial intelligence community. The student will design and develop a toolbox of different algorithms for Bayesian optimization [1,2]. The toolbox will be tested on real problems in computer vision and robot learning.
[1] Jones, D., Schonlau, M., Welch, W., (1998) ``Efficient Global Optimization of Expensive Black-Box Functions''. Journal of Global Optimization} Vol. 13, 455-492.
[2] Ruben Martinez-Cantin, Nando de Freitas, Jose Castellanos and Arnaud Doucet (2007) Active Policy Learning for Robot Planning and Explorationunder Uncertainty. In Proc. of Robotics: Science and Systems.
The student should have knowledge of C/C++.
Semi-supervised learning for humanoids
Intelligent robots have to learn from their environment. However, as small kids, sometimes they need to ask to their parents or teachers to learn more efficiently. For example, the robot might know that there is something interesting, like a person, in the scene that it is seeing. Then, the robot can ask “Is there something interesting in front of me?” or “Is there a person in front of me?”. The objective is to classify the pixels of the image with the label “person” or “not person”. However, the typical answer is: “Yes, there is a person”. The robot has to figure out which pixels corresponds to the person and which not.
Recent developments in machine learning has developed techniques for learning with little supervision. The student will use an available library for that purpose [1]. The robot used for the experiments is the iCub, a new robotic platform developed in the European Union to study the development in infants [2]. The iCub includes a “curiosity” module that can be used to trigger when the question should be done. The main objective of the project is to test the library in a robotic setup and provide the interface with the curiosity module of the robot.
During the project, the student will obtain knowledge in recent research in machine learning, robotics, computer vision and artificial intelligence. Similar techniques are being used by companies such us Google, Willow Garage, Microsoft, Worio...
[1] Peter Carbonetto, Gyuri Dorkò, Cordelia Schmid, Hendrik Kück and and Nando de Freitas. Learning to recognize objects with little supervision. International Journal of Computer Vision, volume 77, May 2008, pages 219-237.
[2] G. Metta, D. Vernon, L. Natale, F. Nori, G. Sandini., "The iCub humanoid robot: an open platform for research in embodied cognition", PerMIS – Performance Metrics for Intelligent Systems Workshop, Washington DC, 19-21 August 2008.
The student should have knowledge of C/C++ and Matlab. Experience in Linux is recommendable.