Problem formulation

Dexterous manipulation in robotics requires learning complex motor skills, which is challenging in open-loop systems due to the lack of feedback mechanisms. Traditional methods, like Dynamic Movement Primitives (DMPs), are effective for encoding movement patterns but often struggle with high-dimensional action spaces. Leveraging latent space representations could address these challenges by simplifying the action space, making it more manageable for open-loop imitation learning.

Task definition

This thesis will focus on developing an open-loop imitation learning framework that integrates Dynamic Movement Primitives (DMPs) with latent space representations for dexterous manipulation tasks. The research will involve designing and implementing algorithms that encode high-dimensional action spaces into more compact latent representations, improving the system's ability to generalize and execute complex manipulation tasks. The performance of this approach will be evaluated through a series of experiments, measuring the effectiveness in terms of accuracy, robustness, and adaptability to different manipulation scenarios.

You shall offer

• Solid knowledge base and experience in deep learning, and robotics.

• Coding skills in Python and C++.

• Experience with ROS.

We will offer

• The most state-of-the-art technologies in deep learning and computer vision.

• Working in a lab with a Germany-wide unique Shadow Teleoperation System.

• Tight support from supervisors, including a workshop on scientific writing.