Towards Robust Grasping with Contact Localization and Grasp Pose Detection von Adrian Marc Zwiener | ISBN 9783843948371

Towards Robust Grasping with Contact Localization and Grasp Pose Detection

von Adrian Marc Zwiener
Buchcover Towards Robust Grasping with Contact Localization and Grasp Pose Detection | Adrian Marc Zwiener | EAN 9783843948371 | ISBN 3-8439-4837-2 | ISBN 978-3-8439-4837-1

Towards Robust Grasping with Contact Localization and Grasp Pose Detection

von Adrian Marc Zwiener
In many fields of robotics and artificial intelligence, algorithms perform equally well as humans. This is entirely different for manipulation tasks like grasping. Grasping is one of the most basic manipulation tasks for robots, interacting with many objects and tools. While humans have no problems grasping a huge variety of objects even without looking, grasping is a challenging task for robots. In this thesis, three sub-modules leading to more robust grasping are discussed.
Firstly, we discuss RGB-D based grasp pose detection (GPD). GPD is one fundamental part in the perception for grasping since it determines where fingers are placed and thus the pose of the manipulator itself. We evaluate two state-of-the-art methods in a scenario where the RGB-D sensor is located on the wrist of the robot.
The second contribution of this thesis is a crosstalk torque calibration with deep neural regression models. In robotic manipulation, small errors can accumulate. Therefore, collisions cannot be neglected. Hence, compliant controllers enhance the safety of the system. Moreover, control laws with collision reflexes can be designed. Without an accurate dynamic model and accurate torque sensing, position and velocity terms have to be introduced into the control law, leading to stiffer joints and a less compliant robot. Crosstalk torques cause inaccurate torque sensing and can severely influence the control of the robot.
The third contribution of this thesis is an extensive discussion and evaluation of different methods for contact point localization with proprioceptive sensors. We are able to efficiently localize contacts on the manipulator’s surface only based on intrinsic sensors, such as position encoders and joint torque sensors. The huge benefit in contact localization methods is that no additional sensors are required. With the help of these localization methods, a tactile feedback for manipulation can be realized.