direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Comparing Deep Reinforcement Learning Methods for Robotic Manipulation



Reinforcement learning describes how agents (e.g. robots) can learn to interact with their environment. Via trial and error, the robot improves its behavior over time. This makes reinforcement learning methods interesting for robotic manipulation, a task that has proven difficult following classic engineering approaches.

Deep learning is a machine learning approach in which models with many layers of abstraction (e.g. neural networks) are learned by backpropagating the error signal through these layers. While the combination of deep learning and reinforcement learning was previously believed to be unstable, a number of recent publications show that this combination can indeed be very powerful.

Description of Work

In this thesis, you will implement different deep reinforcement learning methods and evaluate them in robotic manipulation tasks. To debug the implementation, you will start your work in simulation. After this is working, you will switch to manipulation experiments with a real robotic platform.


Understanding of machine learning, in particular reinforcement learning
Understanding of robotics and control
C++, Python, ROS, Theano/Tensorflow/PyTorch

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions