I am a Graduate student doing my Masters in Robotics at the Robotics Institute in the University of Michigan - Ann Arbor. I am also currently working part time at Honda R&D Americas, LLC as a Student Associate and contributing to their sensor fusion stack.

I am interested in robot perception.

Prior to my studies at Michigan, I worked at Wipro Ltd. in the Robotics CTO team as a Senior Project Engineer. I was mainly involved in building solutions for robot decision making and task planning. I also did a summer internship at Refraction AI, where I worked on lane tracking.

A few of the projects I've worked on so far are listed below:

This project involves coding on an MBot, a mobile robot with differential drive, equiped with magnetic encoders, 2D Lidar and a MEMS 3-axis IMU. The MBot is driven around different maps, and LCM logs are collected. Due to the COVID-19 pandemic, a python based physics simulator was used instead of the actual MBot. The logs along with the simulator was used to perform Occupancy Grid Mapping, MOnte-Carlo Localization, Exploration and Path Planning.

The objective of this project is to program a 6 DOF robot manipulator to perform pick-and-place of colored cubes based on camera readings.

This project comprises of a C based physics simulator to emulate robot movement and collision resolution, path planning using a recursive tree search, and a graphics generator to display the environment. The aim of the program is for a chaser robot to catch the runner robot which performs a random walk in the map environment.

This project compares different Spatio-temporal Neural Network architectures to detect deepfakes on the Celeb-DF dataset. The architectures we are comparing are listed below:

1. Recurrent Convolutional Networks - the pipeline consists of a CNN for feature extraction, LSTM for temporal sequence analysis and Fully Connected Layers for classification

2. R3D - Residual Networks performing 3D convolutions and 3D pooling

3. ResNet Mixed 3D-2D Convolutional Networks - Mixed architecture which starts with 3D convolutions and switches to 2D convolutions in the top layers

4. ResNet (2+1)D - Approximates 3D convolutions by using a 2D convolution followed by a 1D convolution separately

5. I3D - inflates filters and pooling kernels of deep classification ConvNets to 3D, thus allowing spatiotemporal featuresto be learnt using existing successful 2D architectures pre-trained on ImageNet