Research Lead @ mkind · PhD in AI @ Politecnico di Torino
I build learning methods that take robots from understanding 3D scenes to acting on them in real factories: detecting novel and anomalous objects, reasoning about cluttered piles, generating long-horizon motion for surface finishing, and grasping with multiple grippers.
I did my PhD in the VANDAL Lab at Politecnico di Torino and IIT, advised by Prof. Tatiana Tommasi. Today I lead research on Physical AI for industrial surface finishing at mkind in Munich.
One point-cloud model predicts parallel-jaw and vacuum grasps from a shared representation, raising grasp success in real industrial trials.
The first end-to-end model that generates multiple long-horizon robot paths as continuous neural fields, with no post-processing. State of the art on robotic spray painting; patent pending.
Finds the anomalous object in a multi-view scene with about a third fewer parameters and roughly 3× faster training than the previous state of the art.
D3G detects objects and predicts the dependency graph that sets their grasp order in dense clutter. Its benchmark metric was adopted by UNOBench and the ECCV 2026 challenge.
3D-SeND tells known from unknown objects using frozen pre-trained 3D encoders, with no training and only a few reference samples per class. An earlier version appeared at the ECCV 2024 workshops.
Custom CUDA kernels that aggregate superquadric primitives into a voxel occupancy grid, reformulated in log space to remove the numerical instabilities of the original implementation.
C++ library for approximate inference on Bayesian networks. I co-developed the GPU backend: CUDA kernels for logic sampling and likelihood weighting, and OpenCL implementations with a CPU fallback.
Custom simulator and environment for the board game Blokus, used to train reinforcement-learning agents.
© 2026 Paolo Rabino · palettes borrowed from paperqueue