Florian Vogt

Hej! I am a Research Engineer at KTH Royal Institute of Technology in Stockholm, where I develop reinforcement learning algorithms for challenging real-world problems in robotics simulations.

I received my Master's degree from the University of Freiburg. My research lies at the intersection of machine learning and systems, with a focus on making reinforcement learning more sample-efficient, computationally efficient, and scalable.

I am particularly interested in developing simple, efficient algorithms that scale to challenging reinforcement learning problems. This work has led to XQC, a simple yet effective optimization method for off-policy reinforcement learning, whose ideas were later incorporated into FlashSAC, recipient of the RSS 2026 Outstanding Paper Award.

Email / GitHub / Scholar

Research

	FlashSAC: Fast and Stable Off-Policy Reinforcement Learning for High-Dimensional Robot Control Donghu Kim, Youngdoo Lee, Minho Park, Kinam Kim, Aswin Nahrendra, Takuma Seno, Sehee Min, Daniel Palenicek, Florian Vogt, Danica Kragic, Jan Peters, Jaesik Choo, Honglak Lee Robotics: Science and Systems (RSS), 2026 (Outstanding Paper Award) Project Page / Code / ArXiv Optimizing SAC for high-speed robotics training. Adopted as a baseline for high-dimensional robotic benchmarks.
	XQC: Well-conditioned Optimization Accelerates Deep Reinforcement Learning Daniel Palenicek, Florian Vogt, Joe Watson, Ingmar Posner, Jan Peters International Conference on Learning Representations (ICLR), 2026 Project Page / Code / ArXiv Accelerating training by improving the conditioning of the optimization landscape in deep RL.
	Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization Daniel Palenicek, Florian Vogt, Joe Watson, Jan Peters Neural Information Processing Systems (NeurIPS), 2025 ArXiv A study on how normalization stabilizes and scales off-policy learning for complex tasks.
	Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization Sascha Marton, Tim Grams, Florian Vogt, Stefan Luedtke, Christian Bartelt, Heiner Stuckenschmidt International Conference on Learning Representations (ICLR), 2025 (Spotlight) Code / ArXiv Handles information bottleneck issues in tree-structured RL via direct optimization.