Research

Intent Recognition for Socially Aware Navigation

My research into intent recognition for the purposes of Socially Aware Navigation uses a combination of Computer Vision and laser processing techniques to generate observations for use in evaluating the intent of people in a given environment. The goal is to combine the intent recognition with path planning to generate more socially appropriate navigation than traditional path planning. I am using Recurrent Neural Networks to achieve the goal of intent recognition in this research.

Minimalistic Convolutional Neural Networks

Because of my focus on intent recognition for SAN, I have had to use a lot of software aimed at human pose detection. However, much of the systems suffer from slow compute times or scale variant issues. As a consequence of this, I have begun searching for better and faster ways of detecting human pose using CNNs. I am currently working on methods to find and train the minimally sufficient network simultaneously.

Dynamic Action Spaces in Reinforcement Learning

Reinforcement learning is a powerful machine learning technique that allows the system to perform unsupervised learning. However, traditional RL suffers from large state-action spaces. My research in this area focuses on reducing the complexity of RL state-action spaces by performing dynamic allocation of actions in a state space.