Francesco Donato

Robotic manipulation is shaped by a rich collection of geometric structures, from poses, orientations, object shapes, and contact states to tactile observations and task constraints. Rather than treating these quantities as ordinary Euclidean vectors, the research will investigate how differential geometry can provide coordinate-invariant representations, intrinsic distances, geodesic notions of motion, and metric-aware learning principles for systems that must perceive, predict, and act through physical interaction.
The project may focus on several complementary aspects of manipulation, including policy learning, world modeling, planning, control, and tactile-centered perception, with the common aim of using geometric structure as an inductive bias for algorithms that operate on curved, constrained, or learned manifolds. In policy learning, this includes geometry-aware generative models capable of representing multimodal actions, demonstrations, and task-conditioned behaviours while respecting kinematic, dynamic, tactile, and contact constraints; in world modeling, it includes predictive and generative latent models that capture the evolution of robot-object-environment states through ideas such as self-supervised predictive learning in Riemannian latent spaces, geometric flow-based models and diffusion processes, and imagination-based reinforcement learning.
A central question is how to make generated actions, futures, and contact interactions not only statistically likely, but also intrinsically consistent with the geometry and physics of manipulation. The intended outcome is a framework in which differential-geometric principles support learning systems that generalize better from limited data, are more efficient, and produce more physically grounded behaviour in real-world robotic settings, with particular relevance to contact-rich and dexterous manipulation tasks performed with a Franka Panda robotic arm coupled with a dexterous hand or gripper.
This ELLIS collaboration would be a particularly strong fit for the project, as it combines complementary expertise from the Italian Institute of Technology and A141. The Italian Institute of Technology would contribute its experience in complex robotic data collection, with a focus on tactile sensing, tactile perception, task understanding, and the hardware infrastructure required to study contact-rich manipulation, while A141 would provide the mathematical and physics-based expertise needed to formulate and analyze learning algorithms grounded in differential-geometric principles. Together, this collaboration would make it possible to develop novel, robust, and physically grounded algorithms for complex tactile-centered manipulation scenarios.