Aristotle University Of Thessaloniki (AUTH)
Robotic grasping and manipulation of unknown objects in clutter
Abstract
dc:descriptionGeneral purpose robots need the capability to manipulate objects in unstructured environments. Despite the interest of research community, and despite its practical value, robust manipulation of novel objects in cluttered environments still remains a largely unsolved problem. This thesis focuses on two problems: robust grasp planning of unknown objects and manipulating objects in high clutter via pushing actions to facilitate the grasping of an object. Firstly, we consider the problem of planning robust grasps for multi-fingered hands. In contrast to past methods that seek for precision grasps and thus try to achieve force closure with only a few contact points, we plan power grasps since they can compensate for perception errors. To achieve that, we exploit the effect of shape complementarity between the robot hand and the object and demonstrate that this adaptation increases the contact area and thereby the robustness of the grasp. To measure the shape complementarity the unseen part of the scene should be taken into consideration and thus the proposed method performs a 3D shape completion of the hidden surface. The proposed grasp planner demonstrates high grasp success rate both in simulation and real world, even in cluttered scenes. Although this geometric approach produces stable grasps, problems arise in high clutter, due to the lack of collision-free grasp affordances. To deal with the problem of grasping an object in high clutter, we employ pushing actions in order to singulate a target object from its surrounding clutter and create free space for the robotic fingers to perform a prehensile grasp. The proposed methods employ reinforcement learning to train policies that use two pushing primitives for object singulation with the minimum number of actions. In particular, we use the Deep Q-Network (DQN) algorithm and demonstrate that we can learn optimal push policies given depth observations of the scene. To improve learning efficiency, we modify the DQN by splitting the Q-Network (Split-DQN) to two networks, one for each primitive. As results demonstrate, the proposed method leads to faster convergence and increased success rate for the object singulation task and allows the addition of an extra primitive without retraining the existing networks. Then, we propose a modular reinforcement learning method (MRL), which in contrast to previous methods uses continuous actions. This method allows to combine different primitive policies, which may be produced independently by means of reinforcement learning, supervised learning or algorithmic design. Furthermore, it uses effectively prior knowledge by properly shaping the state and action space and achieves singulation in different environments, which traditionally require different manipulation primitives. All the above policies are trained in simulation and are effectively transferred to a real world environment. Finally, we demonstrate that simple heuristic rules that move the target object towards the empty-from-obstacles space can be effective for the singulation task and that guiding the training of the RL agents with these simple heuristic rules in the reward, results to increased success rate. Subsequently, to alleviate the need of multiple pushing actions in order to singulate a object and then grasp it, we propose a push-grasping policy that creates enough space for the fingers to wrap around an object in order to perform a stable power grasp, using a single primitive action. In particular, we employ pushing actions to roll an object into the robot hand and roll out or translate the obstacles from the hand’s closing region in order to create the aforementioned free space. The pushing part of the primitive is followed by closing the robot fingers to establish a power grasp. The power push-grasping policy is trained in a supervised manner by gathering data in simulation. To make the learning process more efficient, the proposed method decouples the action space by learning the hand pose and finger configuration separately. Finally, as the experimental results show, the proposed policy generalizes to novel objects and challenging object configurations and is robustly transferred to a real world scenario.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kiatos, Marios
- Κιάτος, Μάριος
Subjects
dc:subject × 12- Ρομποτική
- Μηχανική μάθηση
- Ρομποτική όραση
- Ενισχυτική μάθηση
- Robotics
- Machine learning
- Computer vision
- Reinforcement learning
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53689