On Open-Source Testbeds for Learning Robotic Manipulation Strategies: Collecting and Comparing Real and Simulated Data

Abstract

Teaching robots to manipulate objects is of critical importance as dexterous, in-hand manipulation is a necessary skill that enables them to operate and interact with a world designed for humans, executing complex tasks. This thesis focuses on the design and development of physical and simulated testbeds that facilitate the collection of data for learning how to execute dexterous robotic manipulation tasks. A modular structure is proposed for the testbed that can be equipped with two different robotic grippers: a fully actuated and an underactuated version. The fingers of both grippers are modular and they can accommodate fingertips with various geometries (e.g., tip curvatures). The testbeds are designed to run unattended so that big data is easily collected and used for training machine learning algorithms for robotic manipulation. The simulated version of the testbed was developed to investigate the differences between simulation and reality and how feasible it is to gather such important manipulation data without needing a physical system. Random Forest and Artificial Neural Networks based machine learning models were trained using both real and simulated manipulation data. These models essentially act as the gripper Jacobians mapping the desired object trajectories to the required motor position trajectories. Three sets of fingertips and two objects were used during data collection and model training to investigate their impact on manipulation performance and to assess the models ability to adapt to different experimental conditions or hardware configurations. It was found that underactuated grippers could adapt to the changes more readily than fully actuated grippers. A series of experiments were conducted to experimentally validate the efficiency of the testbeds and all comparisons and results are discussed in detail. Finally, more work is required to tune the simulation so as to provide data that more closely matches the real data, focusing on phenomena that are difficult to model (e.g., uncontrolled slipping and rolling). All the CAD files of the testbed are distributed in an open-source manner to allow replication by others.