AGGRO: Autonomous Gatherer with Guided Retrieval Operations

Abstract

This thesis presents AGGRO: The Autonomous Gatherer with Guided Retrieval Operations, an innovative robotic system designed to enhance object manipulation in cluttered environments. Building on the foundations of Deep Q-Learning (DQL) and advanced reinforcement learning techniques, AGGRO integrates machine learning, robotics hardware, and sophisticated algorithms to address the "Grasping the Invisible" problem at scale. The system employs a combination of primitive synergies to achieve efficient and precise manipulation of occluded objects. Through comprehensive real-world testing and simulation, the thesis explores various explortation policies, dynamic clutter generation, and the impact of structured clutter scenarios on system performance. The results demonstrate a three policy approach to efficiently reveal targets, fully uncover them, and finally singulate to grasp.