Seminar: Graduate Seminar
Practical Reinforcement Learning for Robust Decision-Making and Vision-Driven Intelligence
Lecturer:
Uri Gadot
Research Areas:
| Our research explores reinforcement learning (RL) as a unified framework for both robust decision-making and intelligent visual systems. The work is organized around two main contributions. The first advances the practical foundations of Robust Markov Decision Processes (RMDPs), introducing scalable and implementable methods to a domain often dominated by theoretical approaches. The second focuses on applying RL to vision-centric tasks, where framing these challenges as sequential decision-making problems enables more effective exploration and improved task performance. The seminar will highlight these applications, including RL-RC-DoT, a task-aware video compression framework that leverages RL to dynamically allocate bitrate to regions of interest, thereby enhancing the accuracy of downstream tasks like object detection and segmentation. Additionally, the presentation will cover FlowRL, an RL-based pipeline optimization method for text-to-image generation within the ComfyUI ecosystem. This approach uses learned reward models to evaluate prompt-workflow combinations without the need for expensive image generation during training, guiding the policy toward generating higher-quality and more diverse outputs. Together, these contributions demonstrate the potential of RL to drive adaptive, efficient, and intelligent behavior in complex visual environments. |
| Uri Gadot is a PhD candidate under the supervision of Prof. Shie Mannor, working at the intersection of reinforcement learning, robust decision-making, and computer vision. His research focuses on developing scalable methods for Robust Markov Decision Processes (RMDPs) and applying reinforcement learning to practical video and image understanding tasks. Uri’s work bridges theoretical foundations with real-world impact, contributing novel algorithms for robust policy learning as well as adaptive systems for video compression and generative image pipelines. Uri has published in leading machine learning and AI conferences such as ICML, CVPR and AAAI. He holds a Bachelor’s degree in Computer Engineering from the Technion.
|