Brett Barkley

I am a PhD student at the University of Texas at Austin advised by David Fridovich-Keil.

I was previously employed by the Johns Hopkins Applied Physics Laboratory, where I worked on AI-based autonomy for aerospace systems.

I have an MS in Aerospace Engineering from the University of Maryland where I was a research assistant under Prof. Derek Paley and member of the Collective Dynamics and Control Laboratory (CDCL). My area of specialization was flight dynamics, stability, and control and my thesis proposed a scalable cooperative autonomy framework for multi-agent aerial reconnaissance.

CV  /  Email  /  Google Scholar  /  LinkedIn

profile photo

Research Interest

I study how to make synthetic data more reliable and effective for scaling AI training pipelines. My work challenges the assumption that even well-structured synthetic data is always beneficial, showing that, without careful integration and diagnostics, it can degrade performance or distort learning dynamics. I develop methods for structured data augmentation, failure analysis, and algorithmic repair to make synthetic data more trustworthy. Recent projects include empirical studies exposing structural flaws in model-based RL pipelines built on synthetic rollouts, time-symmetric data augmentation in sequential decision-making problems, and ongoing development of diagnostic tools for out-of-distribution detection using diffusion models, aimed at identifying when synthetic data distributions diverge from trusted real-world contexts.

News

FTFL teaser
Fixing That Free Lunch: When, Where, and Why Synthetic Data Fails in Model-Based Policy Optimization
Brett Barkley, David Fridovich-Keil
Preprint, 2025
arXiv

We identify when, where, and why synthetic rollouts destabilize MBPO, and propose simple remediations that allow it to perform exceptionally well in environments where it previously could not improve beyond a random policy.

SCOPED teaser
SCOPED: Score–Curvature Out-of-Distribution Proximity Evaluator for Diffusion
Brett Barkley, David Fridovich-Keil
Preprint, 2025
arXiv

We introduce a diffusion-based OOD detection method using score curvature to measure typicality, enabling fast, reliable detection across vision and RL datasets.

STFL teaser
Stealing That Free Lunch: Exposing the Limits of Dyna-Style Reinforcement Learning
Brett Barkley, David Fridovich-Keil
ICML, 2025
arXiv/ Code

We show that synthetic rollouts in model-based RL can actually harm performance, identifying fundamental instabilities in Dyna-style methods.

Time Symmetry teaser
An Investigation of Time Reversal Symmetry in Reinforcement Learning
Brett Barkley, David Fridovich-Keil
L4DC, 2024
arXiv

We explore the role of time symmetry in RL algorithms and propose new augmentation strategies that exploit this structure.


Design and source code adapted from here