The selected candidate will contribute in activities related to the "ROBOEXNOVO - Robots learning about objects from externalized knowledge sources" project funded by the European Union's H2020 Programme with Grant Agreement n. 637076.
Soft robots are at the forefront of AI community research. Differently from their anthropomorphic counterparts, they exploit hundreds to thousands of sensors each, being thus able to generate humongous amount of data that need to be analyzed and understood in order to inform intelligent interactions and behaviors. This poses a wide set of challenging for learning methods called to deal with the modeling of perceptual intelligence of such systems. For instance, an octopus-like soft robot may contain tens of tentacles, each equipped with tens to hundreds of (image) sensors able to (continuously) generate a stream of (~103) pictures/inputs to be ingested and processed. Imaging an applications using a swarm of hundreds of such soft robots quickly raises the number of pictures/inputs to elaborate in the order of 105. Learning a model for representing the environment where the soft robots act requires to process such huge amount of data several times, thus quickly moving the demand for computing resource at the Peta/Exascale level. This in turn, coupled with the challenges posed by a perceiving system so radically different from the current mainstream devices that feed modern computer vision algorithms, calls for a radical rethinking of modern deep vision algorithms.
The goal of our research is to build the algorithmic foundations of intelligent perceptual soft robots. We will investigate how modern deep learning approaches that have proved successful in visual recognition can deal with this completely new setting, both in terms of performance and in terms of computational scalability. To do so, we will design an experimental setting for studying up to 3 scenarios where visual perception is relevant and might support applications of interest for industrial key players.
Strong collaborations with the group of Dr. Barbara Mazzolai is foreseen.
The candidate should have a strong technical and theoretical background, with a PhD in computer science or similar, at least two years of postdoctoral experience with a proved record of research on modern methods of machine learning applied to computer vision and robotics.
Applicants are expected to have a strong publication record in the top journals and conferences in the fields of robotics, computer vision and machine learning.