I am working with the sensors team on the autonomous ID Buzz vehicle for sensor data capture, calibration, visualization, and assessment. I'm currently focussed on developing a ROS based framework for sensor evaluation.

I worked with the Multimedia Video team in developing automation tools in Python for testing video decoding and encoding from HAL through E2E layer. I also worked on developing new tools in the Google Cloud AI platform for the detection of video artifacts using convolutional deep learning models to replace manually executed subjective validations, thus scaling up testing by 80%. For this model, I generated a training dataset of 50k labeled images with visual corruptions emulating decoder and encoder artifacts.

I'm assisting Dr. Armin Sarabi, a research scientist in the Department of Electrical Engineering and Computer Science, in automating the manual process of annotating internet scan data. I've so far worked on analyzing the problem statement and on developing software in Python to create training data for learning models in order to achieve this goal.

During my three months with FIL, I worked independently on developing and testing different features for an AMR in a simulation setup using Gazebo and ROS Melodic framework. I simulated a custom mobile robot in Gazebo by writing URDFs and synthesized data from fake sensors as a close emulation of the real world. I implemented gyrodometry and verified the sensor fusion results using RQT, RViz and other visualization tools. I also implemented features like geo-fencing and interface functionalities. I explored different SLAM techniques (like RTAB, Gmapping and hector) in ROS.

In the second phase, I focused on detection of dynamic obstacles and inclined surfaces in the simulated environment. I analyzed the influence of dynamic objects on gmapping and AMCL, and worked on their detection. I, then, implemented detection of humans in 2D laser data using a random forest classifier. Finally, I got to implement detection of inclined surfaces from 3D point clouds with an accuracy of 0.02 meters in ramp width and of 10 degrees in slope angle estimation, followed by integrating these modules to the mapping and navigation modules.

Working independently, I gained experience in setting my own deadlines to meet the company's expectations, directing research and converting those ideas to plausible implementations. I also had the opportunity to learn new software tools, that I wasn't familiar with but were needed for the work, without letting it affect the deadlines.

Along with 2 other people, I evaluated six theoretical and coding assignments for the EECS 492 Fall 2019 class, consisting of about 150 undergrad and grad students.