Valutazione Empirica della Trasferibilità di Politiche Generaliste di Apprendimento Visuomotorio per la Manipolazione Robotica = An Empirical Evaluation of the Transferability of Generalist Visuomotor Learning Policies for Robotic Manipulation

As embodied AI research continues to advance, the need for adaptable and efficient learning algorithms for robotic tasks becomes increasingly crucial. This Master thesis investigates the performance and versatility of general-purpose models trained on large-scale datasets employed as pre-trained checkpoints and finetuned on different robotic tasks, and compares them against from-scratch robot learning approaches. This research aims to evaluate the effectiveness of pre-trained models in handling various robotic tasks, ranging from simple to complex. The study is structured as a series of experiments comparing different versions of the Octo model and of the Diffusion Policy for visuomotor trajectory learning approach. These experiments are conducted both on a simulated environment as well as on a real Panda robotic arm, and the comparative analysis is performed on identical tasks. We explore the benefits of pretraining on large-scale heterogeneous data and examine how the performance gap between these approaches evolves as task complexity increases. This analysis provides insights into the scalability and adaptability of these models across various robotic applications. We find that robotic vision-action models trained on large-scale heterogeneous datasets are usually not able to properly execute dexterous tasks out-of-the-box on novel environments while finetuning such models does not seem to offer meaningful advantages when compared to approaches requiring training from scratch on tasks of comparable complexity. The thesis also discusses potential improvements to enhance model performance and suggests additional comparative studies that could further our understanding of these approaches. By examining the strengths and limitations of each model, we aim to contribute to the ongoing development of more versatile and efficient robotic learning systems.