Abstract
Many modern robots rely on supervised deep learning models to understand their surroundings, using data from high-dimensional sensors like cameras and lidars. However, training such models require large labeled training datasets, which are difficult and expensive to acquire, especially for novel robot systems or unconventional sensors. In contrast, humans and animals learn perception tasks mostly through self-supervised experience. This talk describes methods that allow a group of robots to learn perception models without external supervision. By performing actions and communicating with each other, the swarm of robots fine-tunes their perception skills, similar to animals honing their motor skills through play. The talk also outlines related research at IDSIA, including finetuning self-supervised perception pipelines deployed aboard 30-gram nano-quadrotors, and applications to the control of exoskeletons.
Bio:
Alessandro Giusti holds a PhD in Computer Science from Politecnico di Milano, 2008, and has 16 years of research experience in Robotics, applied Deep Learning, Computer Vision and Data Visualization. He has been with the Dalle Molle Institute for Artificial Intelligence (IDSIA, USI-SUPSI) in Lugano, Switzerland since 2009, where he is now Full Professor in AI for Autonomous Robotics, coordinating a team of 10 researchers. He took part in a dozen projects in applied research focusing on Deep Learning applications to mobile and industrial robotics; he is the author of more than 80 peer-reviewed publications in top conferences and journals, and the recipient of several awards, most of which for innovative applications of deep learning to various fields.