University of Augsburg · Organic Computing
Research associate and PhD candidate at the Chair of Organic Computing.
I research reinforcement learning and scene graph-based visual reasoning, with a focus on robust, interpretable models and reproducible results.
Open to roles in AI and machine learning
What I'm working on
Scroll or drag to rotate
I work on an assistive robot for people with physical disabilities, built on the Unitree Go2 EDU. It should keep gaining new skills without forgetting the old ones, and learn on its own which skill is the right one in which situation.
In simulation
Proof of concept for learned skill selection: the robot has a fixed library of walking skills and learns from its own experience which one works best on tile, wood or carpet, or when carrying a load. The question: does a personal, continually learning selector beat a static ranking?
On the real robot
I am developing a skill set for the real Go2 so that it can move through buildings on its own, including a follow mode.
3D model: Unitree Robotics (BSD-3-Clause) via MuJoCo Menagerie
Publications, thesis work and open-source contributions.
Preprint · first author · arXiv 2026
Roman Küble, Marco Hüller, Mrunmai Phatak, Rainer Lienhart, Jörg Hähner
How can an agent with a limited action budget build a complete semantic scene graph? The work modernises the navigation component with an updated optimisation algorithm and reworked, factorised action spaces, and evaluates curriculum learning and depth-based collision supervision.
Full paper · second author · ARCS 2025 · Springer
Empirical comparison of the XCS classifier system with established RL methods across 51 tasks from five benchmarks: RL is markedly more reliable in complex multi-step settings, while XCS only keeps pace on simpler, noisy problems.
Vision paper · second author · CoopIS 2025 · Springer
A lightweight MLOC redesign with RL agents in layer 1 and a retunable simulator in layer 2. An anomaly detector separates model drift from agent uncertainty, enabling safe adaptation in dynamic environments.
Master's thesis · University of Augsburg
Modernised an ESSGG agent in Ai2-THOR, replacing the RNN/REINFORCE baseline with a Transformer for long-horizon context and an Advantage Actor-Critic controller. The stable A2C learner drives the gains in graph quality, exploration efficiency and generalisation.
Open source · Julia
Extended the Julia port of MeshGraphNets with native graph objects, custom plotting utilities and a Pluto.jl notebook showcasing the new workflow.
2025 –
Research associate (PhD)
Universität Augsburg · Organic Computing
2024 – 2025
Student research assistant
Universität Augsburg · Organic Computing
2021 – 2024
Working student
Hilti Entwicklungsgesellschaft mbH
2020 – 2021
Internship
Hilti Entwicklungsgesellschaft mbH
2018 – 2020
Working student
Boehringer Ingelheim Pharma GmbH & Co. KG
2023 – 2025
M.Sc. Engineering Informatics
University of Augsburg
Thesis on deep reinforcement learning for embodied scene graph generation
2018 – 2023
B.Sc. Mechanical Engineering
Augsburg University of Applied Sciences
Thesis on load spectrum clustering
Interested in collaborating, exchanging ideas or a position? I'd love to hear from you.