Shubham Singh

Shubham Singh

Robot learning · Manipulation · Human-robot interaction

I'm interested in how robots learn to act in the physical world and how they work alongside the people who instruct them. Most of what I've built comes back to one question: does a learned system keep behaving sensibly when its inputs drift away from what it was trained on? In robotics that question matters most, because a failure is physical.

My current work is an independent study of vision-language-action models, asking whether a robot policy does the same thing when the same request is worded differently. Over the longer run I want to be able to build a robot system end to end, from perception and learning through to control and kinematics.

Before this I spent several years as an engineer on machine learning systems and computer vision software. I hold an M.S. in Computer Science from the University of Texas at Dallas and a B.Tech in Electronics and Communication Engineering from NIT Nagaland, and I live in Dallas.

Interests

Research

Instruction-phrasing brittleness in vision-language-action models

Independent study · ongoing

Vision-language-action policies such as OpenVLA are built on pretrained vision-language models that handle paraphrase well. "Put the spoon on the towel" and "place the spoon on the towel" name the same task, so a policy that kept its backbone's language ability should act the same way for both. I test whether it does. The scene, camera and robot state are held fixed in SimplerEnv and only the instruction text changes, so any difference in behaviour can be traced to wording and not to perception or control.

The scene is held fixed and only the instruction changes. The policy is measured per frame and over full episodes.
Each task has a canonical instruction and twelve paraphrases in six categories. A same-frame probe measures how much single actions change, and full rollouts measure success and trajectory divergence against a control run.

Sensor synchronization for virtual-human interaction studies

Research Assistant, University of Texas at Dallas · 2023

Built a Python service that synchronized and aligned live physiological sensor streams for experiments on how people interact with virtual humans, so researchers could pull every processed signal from one endpoint. Also fine-tuned DistilBERT for multi-label classification and extractive question answering.

Projects

BetaOthello

With Narendra Gangwani · UT Dallas, 2023

An Othello agent whose value function is a Bayesian neural network, trained from self-play with variational inference. Every value the agent computes is a sample from a posterior over value functions, which opens the door to exploring where the agent is least certain.

Image inpainting with deep learning

With Jiazheng Liu and Marin Budic · UT Dallas, 2023

Compared three ways of filling in missing regions of face images on CelebA and CelebA-HQ. A two-stage coarse-to-fine network with contextual attention borrows features from the visible background and gave the sharpest results, a context encoder recovered the overall layout but blurred detail, and a plain GAN served as a baseline. Partial occlusion of this kind is routine in cluttered manipulation scenes.

Experience

Education

Contact

I'm happy to hear from anyone working on robot learning, manipulation or human-robot interaction. Email is the best way to reach me.