Tristan Tomilin - Benchmarking Pixel-Based RL in Egocentric Perception Environments

Описание: UoE RL Reading Group | 17 October 2024

Speaker: Tristan Tomilin (Technical University of Eindhoven)

Title: Benchmarking Pixel-Based RL in Egocentric Perception Environments.

Abstract: In the pursuit of advancing autonomous systems through Reinforcement Learning (RL), benchmarking is essential for assessing performance, validating the ability to interpret and navigate complex environments, and providing a standardized framework for comparison and improvement. However, existing benchmarks for pixel-based learning within embodied environments often face challenges, including high computational demands, a complex setup, insufficient documentation, an absence of standardized metrics, and a lack of baseline evaluations. In this talk, I will introduce new benchmarks targeting these deficiencies, specifically designed for generalization, continual learning, and safe RL.

Link(s): https://ieeexplore.ieee.org/document/...

Bio: Tristan Tomilin is a PhD student in the Data Mining group in the Department of Mathematics & Computer Science at the Technical University of Eindhoven since 2021. With interests spanning across several domains of Reinforcement Learning, including continual learning, multi-agent systems, safe RL, and generalization, his research is characterized by a focus on developing more robust, reliable, and meaningful simulation environments.