This page is still under development.

Traffic Signs Classifier

As part of a Udacity course on Deep Learning, I build a traffic sign classifier using a Convolution Neural Network. This is based on the LeNet architecture. The report on this project can be accessed here.

Learning from Failed Demonstrations

As part of my robotics course at UNC, I implemented the algorithm for the paper "Donut as I do: Learning from failed demonstrations" by Grollman and Billard. I implemented the algorithm in ROS using Python. Using the joint distribution of orientation and angular velocities in failed distributions, the algorithm finds new velocities for successful task performance. The authors had demonstrated the  algorithm on simple scenarios like flipping up a box. We wanted to know if the algorithm would work for a more complex scenario like inverted pendulum. For this a simulation was developed in gazebo to test the performance of the algorithm.  Although the algorithm failed to generalize to this more complex task, certain insights were gained into how it could perhaps be made to work.
Report

---