Pavel Tokmakov

Toyota Research Institute, Carnegie Mellon University

Papers

Total Citations

H-Index

About

Pavel Tokmakov is a researcher whose work sits at the intersection of computer vision, robot perception, and scene understanding. His research spans trajectory forecasting, object tracking, video segmentation, and novel view synthesis — areas critical to enabling autonomous systems to interpret and navigate dynamic environments. One of Tokmakov's most recognized contributions is his work on heterogeneous-agent trajectory forecasting, which tackles the compounded uncertainty arising from both the multiplicity of possible futures and ambiguity in agent state estimation — a nuanced challenge with direct implications for safe robot navigation. This paper has garnered 38 citations, reflecting its relevance to the autonomous driving and robotics communities. His earlier work on motion-based segmentation, exploring how to detect and segment moving objects independent of category, laid important groundwork for open-world scene understanding. More recently, Tokmakov has pushed toward leveraging large pre-trained models for zero-shot open-vocabulary tracking, enabling robots to identify and follow objects without category-specific training — a significant step toward generalizable perception systems. His work on generative monocular dynamic novel view synthesis further demonstrates his range, bridging generative modeling with 3D scene reconstruction. Across his portfolio, Tokmakov consistently advances perception systems that are robust, flexible, and deployable in real-world robotic contexts.