Varun Velpula

Machine Learning Engineer

Third Insight

Biography

Hi! I am a Machine Learning Engineer at Third Insight. Here, I work on a number of computer vision problems for various intelligent systems on government projects from the Navy, the Airforce and DARPA.

I have graduated with my Masters of Science in Robotics from the University of Pennsylvania. Previously, I had done my undergraduate studies at the National Institute of Technology Karnataka in India. My areas of interests are in software development and in the application of machine learning in robotics and self-driving space through computer vision and reinforcement learning. In the past, I had worked on projects in the domains of manipulator, aerial and ground robotics.

During my free time, I bike, read books, draw and listen to music. I particularly enjoy reading non-fiction books, relating to anthropology, world history and current affairs. The weekends are for biking and for catching up with the latest developments in the AI space. I am a massive F1 and endurance racing fan. My favorite genre of music is Indie Rock/Pop and Garage Rock. My favorite artists/bands are The Strokes, Arctic Monkeys, and the Killers.

Interests

Machine learning
Computer Vision
Deep Learning
Artificial Intelligence
Path Planning
Motion Plannning
Localization and Mapping

Education

MS in Robotics, 2021

University of Pennsylvania
BTech in Mechanical Engineering, 2019

National Institute of Technology Karnataka, India

Experience

Machine Learning Engineer

Third Insight

Jul 2021 – Present Austin, Texas

Created a lightweight detection and tracking algorithm in C++ to identify and track objects with robust state tracking in real time on embedded devices (NVIDIA Xavier/NX) with keypoint based optical flow tracker and deep learning tracker using OSNet and YOLO, which were built with ONNX/TensorRT
Created efficient image processing pipelines for generating synthetic datasets with photorealistic 3D models using Blender and OpenCV. Allowed us to create millions of synthetic images with segmentation and ground-truth bounding box label data to feed into our training pipelines for our vision networks to pre-label vision datasets for data engineers to modify
Built company’s machine learning training pipeline and storage / deployment solution utilizing MLOps software frameworks such as MLFlow and DVC. This allowed for an effective evaluation of multiple training experiments and deployment of models into our autonomy stack
Collaborated with Data Engineers and setup out the company’s annotation pipeline, integrating Auto ML solutions for automatic labeling of various internal datasets
Created photorealistic simulation environments with Unreal Engine and utilized Airsim to test developed machine learning algorithms on drones

Teaching Assistant

Department of Computer Science, University of Pennsylvania

Jan 2021 – May 2021 Philadelphia, Pennsylvaia

Formulated course material and assignments for Computer Vision and Reinforcement Learning sections of the course
Taught and guided undergrad and grad students through their classes and course capstone project

Summer Research Intern

ModLab, University of Pennsylvania

Jun 2020 – Aug 2020 Philadelphia, Pennsylvaia

Explored the application of Reinforcement Learning (RL) for policy search for grasping objects using quadcopters. A simulation environemnt was developed on Copellia Robotics (V-Rep) physics simulator, where I designed a custom gripper-quadcopter contraption. The RL algorithms used were value iteration and Q-learning, and they were used to learn the waypoints for motion, yaw angle for optimal grasping orientation and the required angle of the gripper for a successful grasping action.

Research Assistant

ModLab, University of Pennsylvania

Oct 2019 – Dec 2019 Philadelphia, Pennsylvaia

I worked on the ModQuad project under the guidance of Prof. Mark Yim in my first semester at UPenn. The problem statement was to develop a decentralized control algorithm for multi-quadcopter systems to work together in a collaborative swarm for the self-assembly of multi-quadcopter systems during mid-flight.

To achieve modularity, the quadcopters sit inside modular cases. The cases were designed on SolidWorks modelling platform. An important chracteristic of the module structure is that they should be very light (4-5 grams). To acheive this degree of lightness, carbon fiber rods of 1mm are used for the construction. Multiple designs were designed, fabricated and tested. The fabrication of the connectors, used at joints of the structure was done using 3D printing.

Research Assistant

Autonomous Vehicles Lab, Indian Institute of Science

May 2018 – Jul 2018 Bangalore, India

I developed an experimental setup consisting of autonomous Micro Aerial Vehicles (MAVs) to test laws of guidance such as Pure Pursuit and Good Helmsman and tracking laws based on potential vector fields. I implemented the controls using PID tuners for navigation of MAVs in indoor environment using data from IMU sensor and array of motion capture cameras. I had further implemented sensor fusion using Kalman filter in order to minimize noise and improve data reliability. I had developed the network for communication of data among MAVs, motion capture system and workstation using Robot Operating System (ROS) and Matlab.

Engineering Intern

ABB Robotics, India

May 2017 – Jul 2017 Bangalore, India

Integrated an anthropomorphic robotic arm gripper with an ABB industrial robot IRB 1600 to explore its capabilities as a personal robot. Potential applications like typing, writing and pick place operations of irregular shaped objects were explored through experiments. I had designed a gripper based on the open source design by sculptor and designer, Gael Langevin. This design incorporated the under actuation principle to achieve the human-like finger movements. The gripper was interfaced with the robot using an Arduino microcontroller and ABB robot controller IRC-5 to control the gripper finger movements which were using servo motors.