I am an M.Sc. student in Robotics, Cognition, and Intelligence at the Technical University of Munich and a research intern at Agile Robots in Munich. I study 3D object understanding and generative world models for robot perception and interaction.
My current work studies interactive world models for articulated object understanding and manipulation with Xi Wang and Prof. Daniel Cremers, and contact-aware video world models for dexterous manipulation.
I am seeking Ph.D. opportunities starting in 2027 in spatial intelligence, interactive world models, and robot learning for manipulation.
Previously, I collaborated with Prof. Slobodan Ilic, Prof. Benjamin Busam, and Prof. Alois Knoll on 3D object perception and robot manipulation. I earned my B.Eng. from Tongji University.
🎉 Pose Anything Anywhere was accepted to ECCV 2026.
🔬 Started a research internship at Agile Robots, working on world models for robotics and dexterous manipulation.
🎉 VBVR, a large-scale video reasoning suite, was accepted to ICML 2026.
🎉 FUNCanon was accepted to ICRA 2026.
Built foundations in visual localization, 3D object assets, and functional canonicalization for robotic manipulation.
Studying open-world object representations and contact-aware video world models, including articulated interaction in explorable scenes.
Connecting object-centric perception and generative world models to controllable physical interaction.
* equal contribution, † corresponding author
PANY estimates the 6D pose of unseen objects from one or sparse arbitrary RGB or RGB-D references. Its multi-view geometry backbone and pose-graph registration aggregate unposed assist views to improve robustness under occlusion and wide viewpoint changes.
FUNCanon decomposes long-horizon manipulation into actor–verb–object action chunks. Affordance-guided functional canonicalization aligns objects into shared functional frames, allowing pose-aware action primitives to transfer across object categories and tasks.
VBVR introduces a large-scale video reasoning suite with 200 curated tasks and over one million clips. Its rule-based, human-aligned benchmark provides verifiable evaluation of temporal, spatial, and causal reasoning in video models.
TRESP Lab
Research Intern · WZL, RWTH Aachen University
Developed laser spot localization and supported multi-sensor axis-error detection.
Intern · SAIC VOLKSWAGEN
Built ORB-SLAM-based indoor perception and Ackermann lane-following; our team placed second in the AutoPro Smart Driving Competition.
M.Sc. Robotics, Cognition, Intelligence
M.Sc. Robotics, Cognition, Intelligence
Munich, Germany · Master’s thesis: Interactive World Models for Understanding and Manipulating Articulated Objects (ongoing).
B.Eng. Mechatronics
B.Eng. Mechatronics
Bachelor thesis completed at WZL, RWTH Aachen University.
Universal pose conventions and shape representations across open-vocabulary object categories, backed by a scalable canonicalization and annotation pipeline.
The project aims to bring uncategorized 3D models, category-level pose datasets, object-centric datasets, and scene datasets into a consistent canonical coordinate space. I worked on 3D asset preparation and the annotation pipeline for the large-scale canonicalized dataset at TUM CAMP.
Highlight-aware video captioning · Guided Research at TRESP Lab, LMU Munich
Developed a GRPO training framework that rewards video captions for temporal alignment, linguistic fidelity, and brevity. The temporal reward combines moment retrieval and highlight detection signals from UVCOM/Lighthouse-style predictors; the report compares clause-level reward aggregation and DAPO on QVHighlights, with preliminary TVSum experiments.
Mobile Robotics Praktikum · Smart Robotics Lab (SRL), TUM
At TUM’s Smart Robotics Lab, I developed visual-inertial SLAM and trajectory-planning components in C++, then led ROS 1 sim-to-real deployment through control tuning, hardware integration, and mission testing. Our team was the only one to transfer autonomous planning from simulation to real flights; planner tuning and VIO re-initialization raised mission success from 44% to 62% and reduced crashes by 30%.