The thesis builds upon the insight that gaze fixation is useful to structure 3D perception on robots. It explores how adaptability can be enhanced through the integration of Reinforcement Learning for learning optimal policies in various robotic tasks.
The thumb is considered the most important finger for handling objects because it has an exceptional range of motion and can exert various forces. This thesis will investigate and develop various two-thumb designs to improve the manipulation capabilities of the RBO Hand 3. If you are interested in a thesis that involves both algorithmic work and hands-on building, this thesis may be a good fit for you.
The mind's toolbox is a paradigm that explains the flexibility and robustness of problem solving in biological agents. In this project we aim to provide a better understanding of the cognitive and computational mechanisms underlying the strategy selection from the mind's toolbox. Our research has the potential to advance our understanding of how the mind works, with practical implications for a wide range of domains, including psychology, biology and robotics.
Co-design involves jointly programming morphology and control to produce a robot that can robustly solve tasks in a wide range of situations. Automated co-design of hardware and software is a challenge due to the ginormous combined space of morphology and control. To reduce the search space, appropriate inductive biases must be used to guide the search and eliminate portions of the space that are unlikely to contain good solutions.
This work aims to develop a gradient-based control approach that considers a robot's belief to find good trade-offs between task-fulfillment and uncertainty reduction for an example task, tactile localization. The goal is to explore techniques to tailor gradient-based control for tactile measurements by autonomously determining when and where to seek contact to obtain measurements. The work will be implemented in a simulation and transferred to a real-world robotic system.
This thesis explores how agents use information at different levels of abstraction to perform action selection and how they exploit regularities to achieve intelligent behavior in 2D simulation environments. By identifying different levels of abstractions and investigating their dependence on specific tasks, the thesis aims to understand how agents can choose the most effective course of action. This research has the potential to contribute to the development of more advanced and effective artificial intelligence systems.