392271 Robots that learn - manipulation of objects with a human-like robot hand (Pj) (WiSe 2018/2019)

For more information about the project, please contact
Dr. Andrew Melnik anmelnik(at)techfak.uni-bielefeld.de
Office 3.308 (CITEC Building)
Consultation hours: Wednesdays 17:15-18:00

In the project, you will train a human-like robot hand to manipulate objects with unprecedented dexterity in a simulated environment [1, 2] using general-purpose reinforcement learning algorithms [3]. You will experiment with adding touch sensors [4] to the hand and explore generalization performance of learned policies over different objects with and without tactile information. You will compare modern neural network architectures and try to improve on existing approaches. Alternatively, you will explore automatization approaches to break a complex task into a sequence of more simple subtasks to improve generalization performance [5].
[1] http://gym.openai.com/envs/#robotics
[2] https://drive.google.com/open?id=1J2H92AstGpcFYqjmVymMSRKI5xISdB1b
[3] https://github.com/openai/baselines
[4] https://drive.google.com/open?id=1iGIbu00IPmI7IjmJrTUaYLIrKZzcgpdL
[5] https://youtu.be/0og1SaZYtRc

Requirements for participation, required level

- Introduction to Neural Networks course or Advanced Neural Networks
- Python (>= 1 year)