Current students


Section: Computer Science and Engineering

Major Research topic:
Self-supervised learning of general-purpose representations for autonomous robots

For an autonomous robot, the capability to represent the world in a comprehensive and task-independent way is of primary importance.
Current deep learning techniques exploit large amounts of data to extract the most significant features for the particular task they expect to solve.
However, adopting a supervised approach requires enormous amounts of labeled data, data for which additional information has been manually provided to facilitate the learning task. This approach is notably inefficient, as even with such amounts of data, it fails to extract any information from the structure of the data itself.
Moreover, these systems often lead to great performances in specific scenarios, but they fail to scale and produce more general representations of the environment. Indeed, the extracted representations are often useful only for the original task or for very similar ones, where the same data distribution can be assumed.
In robotics, instead, the nature of the observed data can vary widely as the robot moves across different locations in space or as time passes. Dynamic obstacles, adverse environmental conditions, and even the presence of other agents must be well represented by the robot. Finding such powerful representations can further enable active and life-long learning and increase its robustness to unpredictable events.
Current supervised methods seem inappropriate for these objectives. On the other hand, learning good representations without supervision is still an open issue. In recent years, self-supervised learning systems have attracted particular attention, as they provide a framework to exploit the data structure without the need for manual labeling. However, although they show great potential, state-of-the-art supervised systems outperform them.
In this context, my research focuses on finding more efficient techniques to exploit unlabelled data for learning general, robust representations of the environment, which could ultimately improve the autonomy of mobile and service robots.