Josh Roy

I will be starting a PhD in Robotics/AI at Princeton in Fall 2026, advised by Tom Silver. My research interests include abstraction, safety, and continual learning for reinforcement learning, world modeling, and planning for embodied agents — including but not limited to robots.

I am currently a founding engineer at a full-stack AI startup called Bridger, working on LLM, agentic, and machine-learning-enabled automation within professional services including Tax and Accounting.

Previously, I worked as a Machine Learning Engineer on topics ranging from large-scale machine learning to model-based planning at Two Sigma, Cognex, and NVIDIA. I have also been an independent researcher at MLCollective, a non-profit machine learning research lab.

I graduated with my M.S. and B.S. with honors in Computer Science at Brown University, supervised by George Konidaris, Stefanie Tellex, and James Tompkin, in May 2020.

Check out my blog on Medium.

Outside of Computer Science, I teach Taekwondo at Columbia and climb rocks.

news

Sep 27, 2026	Organizing the IROS Workshop on World Models for Robotics.
Aug 03, 2026	Starting my PhD at Princeton.
Apr 12, 2026	Won Bronze at the USA Taekwondo MA State Championships.
Feb 24, 2025	Joined Bridger as the first engineering hire.

selected publications

AAAI
Visual Transfer for Reinforcement Learning via Wasserstein Domain Confusion

Josh Roy and George Konidaris

2021

Abs arXiv Bib

This work addresses the challenge of transferring visual policies in reinforcement learning across domains. We introduce Wasserstein Domain Confusion (WDC) to align visual representations between source and target environments, enabling agents trained in one visual domain to perform well in another without additional training.
@inproceedings{roy2021visual, title = {Visual Transfer for Reinforcement Learning via Wasserstein Domain Confusion}, author = {Roy, Josh and Konidaris, George}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)}, year = {2021}, }
IROS
Advanced Autonomy on a Low-Cost Educational Drone Platform

Luke Eller, Travis Guerin, Botao Huang, and 4 more authors

2019

Best Paper Finalist Abs arXiv Bib

RoboCup Best Paper Award Finalist (3 of 1,127 accepted papers)

Building on the original PiDrone work, this paper demonstrates that sophisticated autonomous behaviors can be achieved on low-cost hardware, making advanced robotics education more accessible. We show that a sub-$300 platform can perform SLAM, autonomous navigation, and multi-robot coordination — capabilities typically reserved for platforms costing 10x more.
@inproceedings{eller2019advanced, title = {Advanced Autonomy on a Low-Cost Educational Drone Platform}, author = {Eller, Luke and Guerin, Travis and Huang, Botao and Warren, Guanyu and Yang, Sophie and Roy, Josh and Tellex, Stefanie}, booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, year = {2019}, }
IROS
PiDrone: An Autonomous Educational Drone using Raspberry Pi and Python

Isaiah Brand^*, Josh Roy^*, Aaron Ray, and 2 more authors

2018

Abs Bib PDF

We created an accessible drone platform that democratizes robotics education by using commodity hardware (Raspberry Pi) and Python, making it affordable for educational institutions worldwide. This work challenged the assumption that meaningful robotics education requires expensive equipment, proving that a sub-$300 platform can teach fundamental concepts from control theory to computer vision.
@inproceedings{brand2018pidrone, title = {PiDrone: An Autonomous Educational Drone using Raspberry Pi and Python}, author = {Brand, Isaiah and Roy, Josh and Ray, Aaron and Oberlin, John and Tellex, Stefanie}, booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, year = {2018}, }