Robot learning · Vision-language-action models · Embodied AI
I study how reliably learned robot policies behave when their inputs shift away from what they were trained on. My current work is on vision-language-action models for manipulation, and the longer goal is to build robot systems end to end, from perception and learning through to control and kinematics.
Instruction-phrasing brittleness in vision-language-action models (independent study, ongoing)
Does a robot policy act the same way when one request is worded differently? I hold the scene, camera and robot state fixed in SimplerEnv and change only the instruction given to OpenVLA, so any change in behaviour can be traced to language. The evaluation harness, paraphrase sets and metrics are public, and the GPU runs are in progress.
A reinforcement learning agent for Othello whose value function is a Bayesian neural network, trained from self-play with variational inference. Built with Narendra Gangwani at UT Dallas.
Image inpainting with deep learning
Three approaches to filling in missing regions of face images on CelebA: a coarse-to-fine network with contextual attention, a context encoder, and a GAN baseline. Built with Jiazheng Liu and Marin Budic at UT Dallas.
Robot manipulation, robustness and evaluation of embodied AI policies, uncertainty in learned control, simulation-based evaluation, and the kinematics and control side of robotics.
M.S. in Computer Science from UT Dallas and B.Tech in Electronics and Communication Engineering from NIT Nagaland.
I work as a software engineer on machine learning systems at Singular Analysts in Dallas. Before that I ported a production computer vision library from Java to C++ for wafer inspection at KLA, and built and deployed machine learning models at PNC.
Python, C++, PyTorch.


