Deep Reinforcement Learning for Physics simulation-based assembly - Studien-/Abschlussarbeit

Kennziffer IPA-2021-69

Studiengang z.B.:
Automatisierungstechnik; Elektrotechnik; Informatik; Kybernetik; Maschinenbau; Mechatronik; Softwareengineering; technische Informatik

Robot programming in assembly automation is a complex and time-consuming task, especially for products in small lot sizes. A novel approach to this problem is the training of robots in modern physics simulations using Machine Learning methods. In particular, techniques of Deep Reinforcement Learning provide promising possibilities for training assembly operations from simulation. In addition, the combination with predefined skills for the force and position control of the robot can significantly shorten the duration of the training. In the context of the present work, the assembly process of electrical cabinet terminals is to be defined for training and then learned. The physics simulation MuJoCo, which is widely used in AI research, serves as the training environment. Through continuous evaluation of the results, variation of the learning algorithms used and subsequent optimization of the learning process, the training is continuously improved.

Scope of the presented work:
Part of the work is the definition of the learning environments (gym environments) for the subsequent use with state-of-the-art learning algorithms (stable baselines). In addition, the simulation environment is prepared accordingly and adapted if necessary. Subsequently, the learning environments are extended by simple approaches of domain randomization. The training results are continuously evaluated and the learning and simulation environments are optimized accordingly. At the end of the work, the assembly process of simple control cabinet terminals should be completely learnable in the simulation and tested with comparable components.

Was Sie mitbringen

- Good programming skills in Python or C++/C
- Experience with Deep Reinforcement Learning an advantage
- Analytical thinking, structured and systematical working method
- Independent and responsible workflow and procedure
- High motivation and personal initiative
- Very good knowledge of German and/or English


Was Sie erwarten können

Tasks:
- Familiarization with the physics simulation and the existing learning framework as well as the learning algorithms
- Definition and implementation of the learning and simulation environments for the training of the respective process phases
- Definition and implementation of domain randomization approaches in the learning environments
- Training and evaluation as well as subsequent optimization of the process execution under variation of the used learning algorithms
(keine Angaben)

Fraunhofer ist die größte Organisation für anwendungsorientierte Forschung in Europa. Unsere Forschungsfelder richten sich nach den Bedürfnissen der Menschen: Gesundheit, Sicherheit, Kommunikation, Mobilität, Energie und Umwelt. Wir sind kreativ, wir gestalten Technik, wir entwerfen Produkte, wir verbessern Verfahren, wir eröffnen neue Wege.

Johanes Kienle
Recruiting
0711 970 1153

Bitte bewerben Sie sich über unser Online-Portal!

Fraunhofer-Institut für Produktionstechnik und Automatisierung
Stuttgart

http://www.ipa.fraunhofer.de

---