Viktor Zhumatiy

Dalle Molle Institute for Artificial Intelligence Research

Papers

Total Citations

H-Index

About

Viktor Zhumatiy is a pioneering researcher in robot reinforcement learning, specializing in overcoming the fundamental challenges of applying machine learning to physical robots: large continuous state spaces, partial observability, and data scarcity. His most influential work, "A robot that reinforcement-learns to identify and memorize important previous observations" (2004, 56 citations), introduced a novel architecture combining reinforcement learning with memory mechanisms, enabling robots to selectively store and recall critical past experiences—a breakthrough for real-world deployment where robots cannot rely on perfect sensors or unlimited trials. Zhumatiy further advanced the field with "Quasi-online reinforcement learning for robots" (2006, 32 citations), where he developed a method that builds probabilistic environment models in real-time while concurrently training policies using anytime algorithms, allowing robots to learn efficiently during active exploration. His work on metric state space reinforcement learning extended nearest-sequence memory algorithms with general trajectory metrics, demonstrating successful vision-based mobile robot control. Zhumatiy’s contributions are particularly notable for addressing the practical constraints of robotics—limited computational resources, noisy sensors, and the need for sample-efficient learning—making his approaches foundational for researchers working on autonomous systems that must learn from real-world interaction rather than simulation.