Alexander Novikov

Papers

Total Citations

H-Index

About

Alexander Novikov’s research lies at the intersection of imitation learning, offline reinforcement learning, and data-driven robotics, with a focus on making robot learning practical without costly online interactions. His most influential work, “Task-Relevant Adversarial Imitation Learning” (22 citations), identifies a critical flaw in adversarial imitation: discriminators often latch onto task-irrelevant visual features, producing uninformative reward signals. This insight has guided subsequent efforts to build more robust reward functions. In “Offline Learning from Demonstrations and Unlabeled Experience” (14 citations), Novikov addresses a key bottleneck in behavior cloning by showing how to leverage mixed-quality, unlabeled data—a common real-world scenario—to improve policy learning without rewards. His “Framework for Data-Driven Robotics” (11 citations) demonstrates a scalable system that uses large logged datasets and learned rewards to accomplish multiple manipulation tasks on a physical robot, bridging the gap between offline algorithms and real-world deployment. With additional work on semi-supervised reward learning for offline RL (7 citations), Novikov’s contributions are shaping how robots can learn efficiently from static datasets, reducing the need for expensive environment interactions. His research is particularly relevant for students and practitioners in robotics, reinforcement learning, and autonomous systems.