Using K-means Analysis and Principal Component Analysis, we attempt to perform feature reduction directly on the observation matrix for POMDPs. The reduced observation matrices allow for faster planning of larger POMDPs, but with minimal loss to the expected reward.
Professor: Prof. Joelle Pineau

While working at BBN Research Labs, I worked on the development of their Byblos speech recognition systems. This involved developing system to recognize languages such as english, arabic, spanish, and mandarin for delivery to a client, as well as experiments for the yearly evaluations.

Gesture recognition is becoming a more common interface for new systems. GT2k attempts to provide researchers and system designers with a toolkit to create the gesture recognition component of larger systems. By using existing recognition technology to provide training and recognition tools, researchers and designers can focus on the more important issues of the system. GT2k builds on top of an existing speech recognition hidden Markov Model toolkit, HTK.
Professor: Prof. Thad Starner

By using simple on-body sensor such as microphones and accelerometers, it is possible to determine the actions of a user. In this experiment, the actions consisted of workshop activities such as hammering or sawing. LDA was used to model the audio data while HMMs were used to model the accelerometer data. By using properities of sound travelling, the audio can be used to aid in parititioning the continuous data, reducing continuous recognition to an isolated recognition problem. The goal of the project was to perform continuous recognition of the users actions, and thus context, based on the simple sensors.
Professor: Prof. Thad Starner

A joint project between NRL, CMU, Swarthmore, and Metrica involved developing GRACE for the AAAI robot competition. GRACE is a RWI B21r robot equipped with VIKIA (a virtual face displayed on an LCD), speech recognition, and speech generation. Our task in the project was to use human robot interaction to get GRACE from the door of the registration hall to the registration desk. This involved locating humans, asking for directions, and acting upon those directions.
Professor: Alan Schultz

The goal of this project is to perform robot localization based on an omnicam mounted to the top of a robot. The omnicam returns a 360 degree view around the robot. Different methods were used to localize. A simplier method involved placing feducials around the building and using a Kalman Filter to maintain position. A more complex method involved applying PCA on the raw images.
Professor: Prof. Frank Dellaert

Using Neural Networks, we are trying to learn low-level control for a physical pinball machine. The pinball machine is interfaced to the computer through a circuit which controls the flippers. A camera tracks the ball in play. This forces us to deal with issues such as real time processing, noise from the environment, and noise from the sensors, as well as learning a complex space.
Professor: Prof. Sven Koenig

As a summer internship at the Naval Research Lab on applying the continuous localization technique on a robot to be used in outdoor environments. Continuous localization is a map-based localization technique initially developed on Nomads. Current work is in porting it to an ATRV Jr, a robot designed for outdoor enviroments. This involved integrating various sensors to help with localization, experimentation in small, high resolution indoor environments, and experimentation in large, low resolution outdoor environments.
Professor: Alan Schultz William Adams

I implemented a gesture recognition system for the perceptive workbench. The perceptive workbench is a desk with a camera beneath which can observe the desk top. By using HMMs (Hidden Markov Models), gestures could be training on the table which could then be used to control applications.
Professor: Prof. Thad Starner