DynaFoRo - AI junior research group

Motivation

Robots have long been indispensable in the German industry and, progressively, in the private sector. They must be increasingly dynamic and able to move at a speed that does not pose a danger to humans. At the same time, they should still be able to perform their tasks precisely and quickly, such as transport logistics. When moving fast and with a high center of gravity or when transporting high and heavy loads, the physical dynamics of the system play a major role in control engineering. It can be modeled manually but is mostly highly abstracted from reality, which is known as the sim-to-real gap. AI and ML approaches can be used to reduce the Sim-to-Real Gap. The robot can learn a suitable model of itself, or it can take over the control entirely through reinforcement learning.

The goal of the interdisciplinary AI junior research group is to use machine learning to optimize the dynamic movement of real robots of different locomotion modes in order to overcome the sim-to-real gap. Two machine learning approaches, hybrid and guided reinforcement learning (RL), will be explored. However, hybrid learning is not intended to replace classical methods but rather to supplement them with machine learning in specific areas where they can actually provide additional performance. Guided RL enables the independent learning of complex control tasks for robots in the real world by integrating existing domain knowledge.

Inspired by the research results of dynamics modeling from the field of humanoid robotics, this project will develop and concretize new ML approaches and platforms for solving current challenges of AI research in robotics and transfer them to different application domains of logistics in order to target practical benefits in logistics as well as in other economic applications.

Future publications within the research group will be listed here.