Paul Misterka

Interested in AI for robotic policies and inference
MSc AI, Robotics @ TU Delft/EPFL

Research

VLA sensitivity to VLM backbonesongoing

How does the choice of Vision Language Model (VLM) impact Vision Language Action (VLA) model capabilities?

Optimizing physics-informed diffusion modelsAug ’26

Optimized PDE-constrained Darcy flow diffusion modelling, cutting training time ~20×. We show that established loss formulations are ill-formed.

VLA robustness in board game scenesApr ’26

A Go-stone placement benchmark showing that state-of-the-art VLA policies collapse to zero success under visual perturbation, because they never ground the coordinate named in the instruction.

Hyperspectral satellite mineral deposit classificationJan ’26

Hyperspectral preprocessing, zero-shot single-image super-resolution and spectral-angle mineral mapping over AVIRIS-NG airborne surveys.

Learning curves for performance- and data-efficient TALJul ’23

Which temporal action localization models hold up when training data, or training time, is the binding constraint.

Benchmarking data efficiency and computational efficiency of temporal action localization models. ICCVW 2023. arXiv:2308.13082

Efficient Temporal Action Localization model development practicesMay ’23

Five-parameter logistic learning curves that predict a TAL model’s full-dataset accuracy from runs on a tenth of the data, and say which architectures are still data-bound.

BSc thesis, Delft University of Technology, 2023 (PDF)

Experience

Projects

Contests