Moritz Schneider

Hi! I am a PhD student at the Neurorobotics Lab at the University of Freiburg, in collaboration with Bosch Corporate Research, focusing on vision-based robot reinforcement learning. I am advised by Prof. Joschka Boedecker. Before starting my PhD, I earned a Master of Science in Automation Engineering and a Bachelor of Science in Business Administration and Mechanical Engineering from RWTH Aachen University. During this time, I worked on disentangled representation learning and interpretability methods for reinforcement learning.

Research

My research aims at discovering principles to build general embodied AI. For this I focus on architectures specifically suited for reinforcement learning (e.g. state-space models), foundations models for embodied AI and pre-training of agents with missing modalities (e.g. without actions).

• Architectures for MBRL: Transformers revolutionized the capabilities in natural language processing and therefore have been the dominant architecture in different machine learning areas. Even though transformers also produce amazing results in reinforcement learning they struggle with long sequences. I am interested in developing specific architectures for model-based reinforcement learning. This includes research on state space models and recurrent neural networks.
• Pre-training of World Models and Agents: Collecting data for embodied agents by hand is tedious. But to autonomously collect large-scale, high-quality datasets, we need good policies in the first place. Therefore, the agent either has to explore the task to infer a good policy by itself or we need to find ways to pre-train agents without action-labeled data like videos. Learning large action-conditioned world models from video to develop a general understanding of the world as well as offline training of agents based on latent actions without cumbersome data collection might therefore be a way to equip agents with a general intuition of actions.

Publications

• The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning
  NeurIPS, 2024, Moritz Schneider, Robert Krug, Narunas Vaskevicius, Luigi Palmieri, Joschka Boedecker
• Interpretable Domain Randomization for Reinforcement Learning with Disentangled Representations
  Master Thesis, 2021, Moritz Schneider
• How Do You Act? An Empirical Study to Understand Behavior of Deep Reinforcement Learning Agents
  arXiv, 2020, Richard Meyes, Moritz Schneider, Tobias Meisen