Transfer learning and limited data learning techniques applied to autonomous robotics

This study reviews recent developments in the field of transferable learning, with direct applicability to autonomous robotics. Through a comparative approach, the paper reviews the conceptual and methodological foundations, highlighting how few-shot and zero-shot learning techniques support the adaptability of robots to novel tasks under data-limited conditions. Relevant applications ‒ ranging from navigation and manipulation to human-robot interaction ‒ are discussed, illustrating the capabilities of these methods. The study also identifies a number of major constraints that prevent the widespread adoption of these techniques and provides recommendations for future research oriented towards transfer scenarios from a simulated environment to reality and between robotic platforms with different kinematics.