frank.roeder [at] tuhh.de
Postdoctoral Researcher - Reinforcement Learning & Robotics

Dr. Frank Röder

I am a member of the Institute for Data Science Foundations at the Hamburg University of Technology, where I research sequential decision-making, world modeling, and causality.

I currently work as a postdoctoral researcher under the supervision of Prof. Nihat Ay on embodied intelligence and teach mathematics, machine learning, and reinforcement learning.

I received my Dr. rer. nat. in reinforcement learning and language grounding in robotics under the supervision of Dr. Manfred Eppe (2021–2025) and Prof. Stefan Wermter (2021–2022).

If you are looking for thesis supervision, please apply via our institute homepage.

Frank Röder

Selected Publications

Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization

Frank Röder, Jan Benad, Manfred Eppe, Pradeep Kr. Banerjee

Conference on Neural Information Processing Systems (NeurIPS) 2025

OpenReview ArXiv Code

We show that dynamics-aligned representations improve zero-shot generalization for contextual world models. DALI integrates within the Dreamer architecture to infer latent context representations from interactions, enabling adaptation to unseen environmental conditions without costly retraining.

Meta-World+: An Improved, Standardized, RL Benchmark

Reginald McLean, Evangelos Chatzaroulas, Luc McCutcheon, Frank Röder, Tianhe Yu, Zhanpeng He, K.R. Zentner, Ryan Julian, J K Terry, Isaac Woungang, Nariman Farsad, Pablo Samuel Castro

Conference on Neural Information Processing Systems (NeurIPS) 2025; also International Conference on Machine Learning (ICML) 2025 Workshop Track CODEML (Spotlight)

OpenReview ArXiv Code

An improved and standardized reinforcement learning benchmark suite for fair evaluation and comparison of multi-task and meta RL algorithms.

Dissertation

Language Grounding in Deep Reinforcement Learning for Dynamic Goal-Oriented Robotics

Frank Röder

Doctoral Thesis, Hamburg University of Technology (TUHH), 2026

Reviewers: Prof. Dr. rer. nat. Nihat Ay, Prof. Dr. Pierre-Alexandre Murena

DOI RL Algorithms Robotics Simulation

Language is the primary medium for instructing embodied agents, yet robots struggle to ground natural language robustly under conversational noise such as disfluencies and polysemy. This thesis studies these limitations in a sparse-reward, language-conditioned reinforcement learning setup, using linguistic feedback and egocentric speech to learn from failure in hindsight — relabeling unintended deviations as alternative goals and predicting goal commands that match the observed behavior. A second pillar tackles action correction, resolving ambiguous or miscommunicated instructions through further verbal input and a context-sensitive hindsight method that models goal uncertainty.

Activities & News

Research Experience

Teaching

Tools and Projects

Music

Industry Experience as Working Student