Safe Object Manipulation Strategies Using Tactile Sensors

Tactile feedback is essential for object manipulation tasks for humans and robots, facilitating reactive control even without vision. This dissertation introduces several tactile-based methodologies to enhance the manipulation performance and safety of robotic systems. Before presenting these methods, the dissertation offers a high-level overview of tactile feedback in human manipulation tasks. Research indicates that humans divide manipulation tasks into distinct action phases, each defined by specific subgoals and guided by tactile cues. This concept is further refined in this thesis through the introduction of three manipulation task stages: grasping, manipulation, and placing. These stages provide a structured framework for categorizing the proposed methods and organization of the thesis. The first approach introduced in this thesis is a novel grasp force controller that is split into control phases, similar to human grasping, and implements several object safety features: It prevents undesired object motions during the grasp by halting finger movements based on tactile cues. Additionally, it maintains a given target force even under external disturbances while holding the object. Subsequently, a reinforcement learning scheme is introduced that enables zero-shot sim-to-real transfer of continuous grasp force control policies. A novel simulation environment is proposed to facilitate this, modeling a grasping scenario with realistically varied objects in size and stiffness. A comparison of these two methods facilitates a broader discussion on the advantages of learning-based methods while also highlighting the beneficial properties of classical controllers. For the manipulation stage, a simulation-based study is detailed, proposing and comparing different tactile sensor configurations for an anthropomorphic robot hand. The performance of each sensor configuration is evaluated by integrating its data into a policy trained on various in-hand manipulation tasks. This comparison of experimental results leads to the formulation of recommendations for developing future end-effector sensorizations. In the fourth and final approach, a supervised learning method for object placing is proposed. Given the challenges of occlusions in vision-based methods in placing scenarios, tactile data proves particularly valuable. The thesis concludes with a comparative analysis of the models, tasks, and sensors employed across the presented approaches.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2024

Author dc:creator

• Lach, Luca Michael

Repository record source_url

https://pub.uni-bielefeld.de/record/2999722

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:2999722