Real-to-Sim Data and Task Generation for LEGO Stacking
ActivePhotorealistic Isaac Sim re-renders of recorded demonstrations let a LEGO stacking policy trained on one 2x2 brick place new bricks, even ones absent …
About Me
I’m a Master’s student in Robotics, Cognition and Intelligence at TUM, interested in general-purpose humanoids and legged robots that operate autonomously in unknown environments. In particular, I’m curious how robots can learn on their own by generating their own training data, supervising themselves, and continually improving from experience, while reusing what they learn across environments and tasks.
News
- Aug 2026 My team won the challenge at the RimA Summer School by FZI and KIT in Karlsruhe.
- Feb 2026 Joined the Autonomous Vehicle Systems Lab at TUM as Research Assistant.
- Dec 2025 One-week hackathon by LSY and Poke and Wiggle: trained diffusion policies from 100+ VR demonstrations and deployed them on a Franka Panda.
- Oct 2025 Joined the Learning Systems and Robotics Lab at TUM as Research & Teaching Assistant.
Featured Projects
Imitation Learning of Optimal Trajectories for Jumping Quadrupeds
Active📢 Stay tuned for future updates!
Data-Driven Diffusion Policies for Robot Imitation Learning
CompletedCollected human demonstrations for a LEGO Duplo stacking task using a VR-setup (Metaverse headset), learned Diffusion Transformer Flow policies based …
Building TUM in a Day
CompletedA transformer-based pipeline for dense and consistent 3D reconstruction from monocular RGB videos of large-scale real-world scenes.
Publications
Learning Where it Matters: Autonomous Data Collection and Offline Reinforcement Learning for Local Contact-Rich Manipulation
arXiv preprint · 2026 Under review
Autonomous real-world data collection focused on the contact-rich part of a task, turned into policies with offline RL and combined with motion planning: 96% success on assembly tasks from only 2–2.5 h of data.
@misc{hausdorfer2026learning,
title = {Learning Where it Matters: Autonomous Data Collection and Offline Reinforcement Learning for Local Contact-Rich Manipulation},
author = {Oliver Hausd{\"o}rfer and Linus Schwarz and Gabor Marko and Christian Dietz and Timo Class and Luka Hofer and Jim Yun-Jin Li and Johannes Hechtl and Ralf R{\"o}mer and Angela P. Schoellig},
year = {2026},
eprint = {2607.15982},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2607.15982}
}Experience
Research Assistant
Autonomous Vehicle Systems Lab (TUM)
2026 - Present
Munich, Germany
Developing an agentic pipeline that turns task prompts into robot programs for autonomous data collection. Set up single- and bimanual robot arms with simulation twins, teleoperation and data collection, fine-tuned and deployed VLAs, and benchmarked optical and IMU-based motion capture for whole-body teleoperation of a Unitree G1.
Research & Teaching Assistant
Learning Systems and Robotics Lab (TUM)
2025 - 2026
Munich, Germany
I built a photorealistic simulation pipeline for co-training robot manipulation policies on synthetic images. Teaching Assistant for the Advanced Robot Learning and Decision-Making course.
Intern, Working Student & Undergraduate Student - Bachelor Thesis
2022 - 2024
Abstatt, Germany
I analyzed and validated terabyte-scale ADAS test drive data, developed, debugged, and virtualized a scalable analysis tool for cloud deployment and designed and maintained CI/CD pipelines for automated testing and deployment. Thesis title: “Fusion of Video Objects and Radar Detections for Automatic High Beam”.
Student Research Assistant
2021 - 2023
Esslingen, Germany
I developed Sensor Fusion, Planning, and Control algorithms in simulation for an autonomous racecar system and supervised and coordinated student projects.
Education
