Ankit Sinha

Hello 👋, thanks for visiting my profile! I'm a Master's student in Computer Science at the Georgia Institute of Technology, currently pursuing robotics research under the guidance of Prof. Sehoon Ha. Prior to this, I worked as a Software Developer at Oracle Cloud Infrastructure. I earned my B.Tech in Computer Science and Engineering from the Indian Institute of Technology (BHU), Varanasi. During my undergraduate studies, I had the opportunity to collaborate closely with Prof. Pratik Chattopadhyay and Prof. Prasenjit Chanak. Feel free to explore my projects and reach out if you'd like to connect!

Email  |  CV  |  LinkedIn  |  Github |  Substack  |  Google Scholar

Research

My current research interest is in Robot Learning. I like to develop robotic controllers using learning-based methods that can involve Sim-to-Real RL, Offline RL, Imitation Learning, etc. However, I come from a Data Science / ML background with experience in Computer Vision and Natural Language Processing (pre-GPT era). Below, you can find some of my ongoing and past projects.

Developing an IsaacLab extension to simulate and optimize the morphology of the BALLU (Buoyancy-Assisted Legged Locomotion Unit) robot with the goal of achieving efficient and dynamic locomotion.

We implement an end-to-end Deep Learning pipeline consisting of a Faster-RCNN-FPN localizer and a ResNet-18 encoder for product recognition in retail stores.

Projects

Here are some of the projects I've worked on, showcasing my skills in Machine Learning, Data Science, and Reinforcement Learning.

Implemented the Trust Region Policy Optimization (TRPO) algorithm in pytorch and validated on standard mujoco continuous control tasks. It is further extended to solve a much challenging problem of Quadrupedal Locomotion using CPG-based gait priors.

Implemented a Deep Q-Network (DQN) agent to play Video Poker and compare its performance against a Dynamic Programming (DP) based optimal agent. This project explores how closely a model-free RL agent can learn the perfect strategy in a complex, yet solvable game like 3-turn Video Poker.

PoC of a pipeline that processes scanned documents from arbitrary images and classifies them into their respective categories with 98-99% accuracy. The system combines Computer Vision (OpenCV, Tesseract) with NLP (Word2Vec, GloVe embeddings) and Deep Learning for robust document understanding.