Congratulations Juan Del Aguila Ferrandis for achieving Master of Research

Reinforcement Learning for Non-prehensile Manipulation under Visual Occlusions and Obstacles

Congratulations to Juan Del Aguila Ferrandis for achieving Master of Research, awarded with distinction, entitled Reinforcement Learning for Non-prehensile Manipulation under Visual Occlusions and Obstacles.

Summary

Manipulation without grasping, known as non-prehensile manipulation, is essential for dexterous robots in contact-rich environments, but presents many challenges relating with underactuation, hybrid-dynamics, and frictional uncertainty. Even simple non-prehensile tasks, such as pushing a box on a table, are difficult to address using either model-based or model-free methods. Model-based methods generally restrict the available contact surfaces to avoid combinatorial explosion and assume quasi-static motion, which limits the object velocity. On the other hand, model-free methods typically restrict the problem scope, for instance by disregarding the target object orientation, since they require vast amounts of interaction data to reduce out-of-distribution scenarios and struggle with long-horizon exploration and credit-assignment. In this thesis, we consider the non-prehensile manipulation problem through a model-free paradigm with reinforcement learning (RL). In particular, due to the long-horizon setting with contact uncertainty, previous non-prehensile works have struggled to handle visual occlusions as well as to incorporate obstacle avoidance in RL control policies, which we aim to address.

Examiners

Mehmet Dogar (external)

Steve Tonneau, University of Edinburgh (internal)