Leonid Keselman

Papers

Total Citations

H-Index

About

Leonid Keselman is a robotics researcher whose work bridges the gap between human intuition and machine learning, with a particular focus on developing algorithms that learn effectively from limited or indirect feedback. His key research areas include human-robot interaction, optimization from user preferences, and perception for robotic manipulation. Keselman’s major contribution lies in rethinking how robots acquire skills: rather than relying on traditional metric scores or ground-truth data, he pioneers methods that learn from pairwise user preferences. This approach is especially valuable in human-centric contexts where objective performance metrics are unavailable or impractical. His 2023 paper on optimizing algorithms from pairwise user preferences (6 citations) demonstrates a paradigm shift toward more intuitive, human-aligned robot learning. In parallel, Keselman has advanced robotic perception through his 2024 work on shape from shading for manipulation (3 citations), where he shows that controlled illumination can efficiently generate high-quality surface normal and depth discontinuity information at low computational cost—a practical solution for tabletop-scale robotic tasks. By making robot learning more accessible to developers without ground-truth data, Keselman is helping democratize advanced robotics capabilities.