Graph Gaussian Processes for Interactive Robot Task Learning

Giovanni Franzese – TU Delft

PhD candidate

The adaptability of robot manipulators to many different tasks is currently constrained by systematic hard coding of each specific task. Recent machine learning methods like Learning from Demonstrations (LfD) and Reinforcement Learning (RL) have shown promising results in having fast reprogramming of the task using human demonstrations or adopting a trial and error strategy. However, these machine learning techniques struggle in generalizing when in novel, previously unseen situations, losing trust for possible applications in industrial and household environments. Recent research aims of making robots aware of what they do not know when facing an uncertain or ambiguous situation and directly asking for human help. Successful results were obtained using Gaussian Process for encoding of nonlinear robot policies thanks to a clear mathematical estimation of the model uncertainty. Nevertheless, a clear formulation of Gaussian Process for graph inputs is still not well studied and understood despite the fact that Graph regression is showing promising results for better interpretability and generalizability in solving complex robotics tasks.

Keywords: Interactive Imitation Learning, Graph Learning, Gaussian Process

Scientific area: Machine Learning for Robotics