Representation Learning for Robotic Control, Multi-Task Goal-Conditioned Policy Learning, Robotics, Embodied AI
Maximilian joined the Intelligent Autonomous Systems Group as a Ph.D. student in November 2022. He is employed as a Researcher at the DFKI Research Department SAIROL under the supervision of Dr. Boris Belousov and Prof. Dr. Jan Peters.
Before starting his Ph.D., Maximilian studied at RWTH Aachen and completed his Bachelor's degree in Mechanical Engineering and his Master's degree in Automation Engineering. At IAS, he wrote his Master's Thesis "Curriculum Adversarial Reinforcement Learning" under the supervision of Prof. Dr. Carlo D'Eramo and Prof. Dr. Georgia Chalvatzaki. In his thesis, he explored a novel combination of concepts to improve the training of robust policies.
- Statistical Machine Learning (SS 2023)
- Robot Learning (WS 2023/2024)
I am inspired by the vision of affordable robots that support us during our everyday life. Communicating tasks to such robots should be as easy as interacting with other humans using our natural language. While today's language models already equip us with common ground knowledge to understand a language instruction, the transfer into low-level robotic actions is a highly complex problem yet to be solved. Current approaches record large datasets of language labeled robotic trajectories and train goal-conditioned policies to imitate recorded behavior. However, even the largest model at present, RT-2, is not able to generalize from its skill distribution during training. Data scaling is the most common solution hypothesis to this problem. I believe that we lag on solutions to more fundamental research problems. What is a good vision-language task representation for robotic control? How do we learn such an action-centric task representation? Transforming such task representations into robotic actions, how do we better represent and learn multimodal action distributions for language-conditioned visuomotor robotic control?
- , CoRL Workshop on Language and Robot Learning: Language as Grounding.