(Posted on Feb 22th, 2011) It is a hard problem to get proper feedback when using Natural Interface to control your PC. KinVi is developed to address this problem by constructing a virtual environment that interact with human. In KinVi, the user can control Windows 7 by interacting with the virtual touch pad, handwriting pad, buttons, doors and other objects. By this means the user can get the most intuitive and natural visual feedback when using NI. After a period of training, the user can even make the 3D world window transparent or invisible and still be able to perform correct gestures.
(Posted on Jan 4th, 2011) This is the demo of our Kinect Win7 control program which provides full support of mouse and keyboard operations. This demo simulates the 8Pen Android input application and shows the advantages and potential applications of using novel text input methods for natural interface against traditional ones, such as an on-screen keyboard. This program uses OpenNI packages.
(Posted on Dec 21st, 2010) Background modeling and multi-human tracking in 3D space with Kinect Sensor.
Transfer Learning (TL) now has three major directions: what to transfer, how to transfer and when to transfer. In order to find an answer to the first question, one must define what knowledge consists of and how to describe common knowledge between different domains and tasks. Several knowledge representation methods have been proposed, but few can be used as a general framework for more than one TL application. The second question (how) is mostly investigated in AI applications. Researchers have applied the transfer learning techniques to some reinforcement learning tasks such as the inverted pendulum and mountain-car tasks. However, they can only provide results in simulation. The third question is concerned with the validity and effectiveness of the knowledge transfer. The transfer is valid only when there are some similarities between source and target. However, most researchers tend to ignore this topic by simply assuming that the source and the target are similar. The focus of this project is to design a complete transfer learning framework for a real-world fully dynamic system, such as a robot.
This framework will be complete because it will not ignore any of the three questions. Intuitively, the three questions are related to each other. For example, the answer to the what question provides the mathematical abstraction of knowledge, decides how the knowledge is transferred and distinguishes common knowledge and the knowledge that is specific for each domain or task. It also decides under what situations (when) only the common knowledge needs to be transferred. This framework will no longer to be limited to simulations, real robot experiments will demonstrate the complete transfer learning procedure.
Develop emergency response behaviors for the MDS (Mobile/Dexterous/ Social) robots for victim search and rescue. The approaches were designed for the uBot-5 at UMass and ported to the Nexi robot at MIT Media Lab.
Please check our project webpage on http://www-robotics.cs.umass.edu/~yunlin/search/.