Aristotle University Of Thessaloniki (AUTH)
Ρομποτική λαβή και χειρισμός αντικειμένων με επαφή με το περιβάλλον
Abstract
dc:descriptionRobotic grasping and manipulation in uncertain environments is still one of the unsolved problems in robotics. This thesis focuses on two problems: robotic grasping with environmental contact and total singulation of a target object in clutter to facilitate its grasping. Firstly, three grasp strategies are proposed which exploit environmental contact for grasping flat objects from support surfaces. The flatter the object (e.g. a smartphone or a plate), the more difficult to find a grasp without colliding with the support surface. Inspired by humans, the proposed robotic grasp strategies utilize compliant contact with the support surface, which leads to robustness against uncertainties that may be introduced by the perception system and cover a wide variety of scenes. Furthermore, a decision-making mechanism is proposed in order to select the proper strategy given the scene. The strategies were implemented in a wide range of robotic hardware and experimentally evaluated in a variety of use cases. In particular, the robotic arm KUKA LWR4+ was used along with multi-fingered hands such as Barrett-Hand BH8-282, Shadow-Hand Lite and Smart Grasping System. Furthermore, the method was implemented on the robotic platform RAMCIP and tested in realistic use cases, taken place in humans' apartments. Subsequently, the thesis proposes two singulation methods preceding the above grasp strategies and facilitating their usage in cluttered environments by creating free space around the target object. Both methods use reinforcement learning to train policies that use two pushing primitives, in order to totally singulate the target in the minimum number of actions. The first proposed method is called SplitDQN and modifies Deep Q-Network (DQN) by splitting the Q-Network to use one network per pushing primitive. In the experimental results is shown that the proposed method leads to increased success rate, faster convergence and allows the addition of a new primitive without retraining the existing networks. The second proposed method is a modular reinforcement learning method (MRL) which, in contrast to SplitDQN, uses continuous actions. A high-level policy is trained to select between primitive policies, which are learned separately. This method allows to combine different primitive policies, which may be produced independently by means of reinforcement learning, supervised learning or algorithmic design. Also, it uses effectively prior knowledge by properly designing the pushing primitives and reducing the variability of the state representations, which leads to increased success rate with respect to state-of-the-art methods. Furthermore, the method achieves singulation in different environments, like a surface with open limits (table) and a surface with walls on its limits (bin), which traditionally require different manipulation methods. The policies that are trained with the above two methods are effectively transferred to a real-world environment in both cases, with MRL outperforming SplitDQN.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sarantopoulos, Iason
- Σαραντόπουλος, Ιάσων
Subjects
dc:subject × 12- Ρομποτική λαβή
- Ενισχυτική μάθηση
- Χειρισμός αντικειμένων
- Robotic grasping
- Reinforcement learning
- Object manipulation
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Ρομποτική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Robotics
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/51395