Advanced Multi-Fingered Robotic Hands: Dextrous Manipulation and Learning AI

What it is about

In this module you will learn how multi-finger robotic grippers with high-resolution tactile sensors and modern, simulation-based learning methods are used to achieve dexterous manipulation close to human capabilities. By the end you will be able to mathematically model hand and sensor physics, interpret fast GPU-based simulations, and understand and evaluate learning-based strategies for grasping, in-hand manipulation, and tactile fine manipulation.

What you will be able to do

• Mathematical modeling of robotic multi-finger hands (kinematics, grasp matrix, impedance control)
• Construction and evaluation of fast physics-based simulations for handling and contacts
• Application of learning-based methods for grasp finding and multi-contact grasp evaluation
• Use of reinforcement learning for dynamic in-hand manipulation and state estimation
• Understanding and use of high-resolution spatio-temporal tactile sensing and data-efficient learning methods
• Methods for robust transfer of learned strategies to real robot systems
• Concepts for autonomous calibration of hand kinematics and sensing
• Combination of whole-body planning and grasp planning for humanoid systems

What the module consists of

• VorlesungVermittlung der Theorie zu Handphysik, Simulation und lernbasierten Methoden mit Beispielen
• GastvorträgePraxisperspektiven aus Industrie (Agile Robots) und zu GPU-/Cloud-Architekturen (NVIDIA)

Teaching method

• Präsentation (Slides oder Tafel)Erklärung der physikalischen Modellbildung und lernenden Methoden; Veranschaulichung an kleinen Beispielen
• Q&A zu Beginn jeder SitzungWiederholung offener Fragen, Besprechung von Beispielaufgaben und deren Lösungen
• LiteraturangabenVerweise auf aktuelle Forschungsarbeiten zur Vertiefung jedes Themas