Romain Guntz
Master's student at ETH Zürich, Soft Robotics Lab — supervised by Prof. Dr. Robert Katzschmann.
01About
Passionate about intelligent decision-making and robotics, I'm exploring physical AI through dexterous manipulation. My research focuses on sample-efficient reinforcement learning applied to flow-based policies — including normalizing flows and few-step flow matching — with the aim of enabling real-world reinforcement learning without relying on simulation.
Concretely, I study how reinforcement learning can improve imitation-learning policies. The RL algorithm I study is sample-efficient: an online approach that injects expert demonstrations into the replay buffer and relies on a well-conditioned critic to accelerate value-function convergence in sparse-reward tasks while avoiding value overestimation.
In parallel, I work on humanoid robots for loco-manipulation. My longer-term goal is to fine-tune large Vision-Language-Action (VLA) models on long-horizon industrial assembly tasks.
Beside research, I play football, read, and produce music as a hobby beat-maker.
02Education
-
University of Oxford — Master's Thesis Jul 2026 – Jan 2027
- Dexterous Manipulation via Flow Policies and Sample-Efficient Reinforcement Learning.
-
ETH Zürich — M.Sc. Robotics, Systems and Control Expected 2027
- Average: 5.18/6.0
- Coursework: Robotics, Control, Computer Vision, and Artificial Intelligence.
-
EPFL — B.Sc. Mechanical Engineering Sep 2021 – May 2024
- Average: 5.15/6.0
- Coursework: Dynamics and Control Systems, Fluid Mechanics, Solid Mechanics, Material Science.
03Experience
-
ETH Robotics Club — AI Research Engineer Apr 2026 – Present
Humanoid loco-manipulation via Vision-Language-Action models and reinforcement learning fine-tuning.
-
BMW — Robotics Research Engineer (Internship) Sep 2025 – Mar 2026
Researched demonstration-guided reinforcement learning to fine-tune robotic foundation models.
04Projects
More on GitHub.
-
Soft Robotics Lab
Improving the SERNF pipeline, a sample-efficient off-policy reinforcement learning fine-tuning framework leveraging normalizing flows and critic chunking.
-
SCAI Lab
Instruction-following indoor navigation pipeline using text-conditioned topological maps, cross-modal attention, 3D feature fields, and waypoint prediction from a monocular frontal camera.
-
RSL Lab
Reduced inference time of the Gaussian Splatting library gsplat by 23% using pruning techniques.
05Personal Projects
-
Backgammon Strategy Solver 2023
Developed a Monte Carlo Tree Search (MCTS) simulation in C to solve Backgammon games.
-
Designed a deep learning model to predict the outcomes of Greyhound racing events.
06Reading
- Le Tour du malheur (I, II) · Le Lion
- Dora Bruder · Accident nocturne · Des inconnues
- The Old Man and the Sea · The Sun Also Rises · A Moveable Feast · The Garden of Eden
- Dubliners
- Le Mur · Les Mots
- L'Étranger · La Peste
- The Minority Report · War Game · We Can Remember It for You Wholesale (Total Recall)
- L'Oiseau bleu d'Erzeroum
- La Vérité sur l'affaire Harry Quebert (The Truth About the Harry Quebert Affair)
- L'Écume des jours
- The New York Trilogy