From Constraints to Opportunities: Efficient Object Detection Learning for Humanoid Robots

Reliable perception and efficient adaptation to novel conditions are priority skills for robots that function in ever-changing environments. Indeed, autonomously operating in real world scenarios raises the need of identifying different context’s states and act accordingly. Moreover, the requested tasks might not be known a-priori, requiring the system to update on-line. Robotic platforms allow to gather various types of perceptual information due to the multiple sensory modalities they are provided with. Nonetheless, latest results in computer vision motivate a particular interest in visual perception. Specifically, in this thesis, I mainly focused on the object detection task since it can be at the basis of more sophisticated capabilities. The vast advancements in latest computer vision research, brought by deep learning methods, are appealing in a robotic setting. However, their adoption in applied domains is not straightforward since adapting them to new tasks is strongly demanding in terms of annotated data, optimization time and computational resources. These requirements do not generally meet current robotics constraints. Nevertheless, robotic platforms and especially humanoids present opportunities that can be exploited. The sensors they are provided with represent precious sources of additional information. Moreover, their embodiment in the workspace and their motion capabilities allow for a natural interaction with the environment. Motivated by these considerations, in this Ph.D project, I mainly aimed at devising and developing solutions able to integrate the worlds of computer vision and robotics, by focusing on the task of object detection. Specifically, I dedicated a large amount of effort in alleviating state-of-the-art methods requirements in terms of annotated data and training time, preserving their accuracy by exploiting robotics opportunity.