Tai Hoang
PhD Student in Machine Learning and Robotics at Karlsruhe Institute of Technology.
4th Floor InformatiKOM 1.
Adenauerring 12, Karlsruhe
I’m a PhD student at the Autonomous Learning Robot (ALR) group at KIT, advised by Prof. Gerhard Neumann. Previously, I was a research assistant at the Volkswagen Machine Learning Research Lab, working with Dr. Maximilian Karl and Prof. Patrick van der Smagt on world models and model-based reinforcement learning. I hold an M.Sc. from TU Munich and a B.Sc. from the University of Information Technology, Vietnam.
research
I develop principled methods for effective information propagation across space and time in physical modeling and robot learning. Inspired by principles from dynamical systems, I design graph-based simulators to predict deformable dynamics and reinforcement learning policies for robotic control.
Research interests: reinforcement learning, geometric deep learning, graph-based simulation, and physical world models.
Selected projects:
Also in the lab: meta-learned graph simulators (MaNGO, NeurIPS 2025), diffusion policies for massively parallel RL (ICML 2026), and world models for model-based RL. Full list on the publications page.
news
| Sep 26, 2026 | Our paper Long-Range Spatio-Temporal Graph Propagation Through Oscillations is accepted at NeurIPS 2026! |
|---|---|
| Jun 02, 2026 | I gave an invited online talk hosted by Sergio Valcarcel Macua at Microsoft Research, Cambridge on Graph Neural Modeling for Deformable Manipulation — with Geometry and Physics as Inductive Biases, covering our recent works IGNS and HEPi. |
| May 01, 2026 | Two papers accepted at ICML 2026: Trust-Region Diffusion Policies for massively parallel on-policy RL, and PAWS (preference learning with advantage-weighted segments). |
| Jan 22, 2026 | IGNS is accepted at ICLR 2026! We improve long-range interactions in graph neural simulators via Hamiltonian dynamics. See the project page. |
| Sep 18, 2025 | Two papers accepted at NeurIPS 2025: MaNGO (adaptable graph network simulators via meta-learning) and AMBER (adaptive mesh generation). |
