Grounding and guidance in VLA policies. Unlearning in diffusion models.
First author · IJCB 2026
PIU protects a chosen identity in ID-conditioned diffusion models by re-teaching what its identity conditioning means, so the model stops generating that person while image quality stays intact.
Computer Vision Laboratory, University of Ljubljana
Thesis work · 2026
Recover expert perceptual structure from frozen X-VLA representations, including DINOv2-like tokens and CeDiRNet-aligned maps, to support action generation and representation alignment.
Visual Cognitive Systems Laboratory (ViCoS), University of Ljubljana
Research project · 2026
Language-conditioned queries score the native X-VLA image-token grid to reveal which visual regions matter for the current manipulation step.
Visual Cognitive Systems Laboratory (ViCoS), University of Ljubljana