Postdoc, Learning & Adaptive Systems Group, ETH Zürich
I’m a postdoc in the Learning & Adaptive Systems Group at ETH Zürich, supervised by Prof. Andreas Krause, since September 2025. My research aims to enable autonomous systems, physical or virtual, to operate in the real world through large-scale multimodal alignment, uncertainty-aware adaptation, and human feedback.
Before joining ETH, I was a Ph.D. student at RPL (Robotics, Perception and Learning), KTH Royal Institute of Technology, and a research visitor at Yale University in the Interactive Machines Group. I am also a member of the WASP doctoral program.
Previously, I researched combinatorial optimization using machine learning at TU Delft in the Air Transport and Operations department, and interned at Latent Logic, a self-driving car startup later acquired by Waymo.
| Jun 2026 | MOSAIC published at ICRA 2026, reframing preference-based RL as multi-objective learning from language explanations. Project page. |
|---|---|
| Jun 2026 | Uncertainty Quantification for Flow-Based Vision-Language-Action Models accepted to the RSS 2026 Diff4RL Workshop, introducing VFD for failure detection and SAVE for active VLA fine-tuning. Project page. |
| May 2026 | Awarded a compute grant from Swiss AI for CSCS Alps extreme-scale HPC access. |
| Apr 2026 | Reinforcement Learning via Self-Distillation accepted to ICML 2026. ICML page · website · code. |
| Mar 2026 | SDPO accepted to the ICLR 2026 Workshop on Scaling Post-training for LLMs (SPOT). |
| Sep 2025 | Joined the Learning & Adaptive Systems Group at ETH Zürich as a postdoc. |
| May 2025 | FLoRA published at ICRA 2025, adapting reward functions with low-rank style parameters for sample-efficient preference-based RL. Project page. |
| Nov 2024 | Awarded a research visit to Japan focused on reinforcement learning and industrial AI applications for manufacturing, including automotive manufacturing with Toyota Motor and Toyota Industries Corporation, manufacturing and computer vision with Toshiba, robotics with Hitachi, and AI/finance with DMG Mori, with the Swedish Embassy in Tokyo and Nagoya. |
| 2024 | POLITE was a finalist for the ICRA 2024 Best Conference Paper Award, Best Student Paper Award, and Best Paper Award on Human-Robot Interaction. |
| Apr 2024 | Started a research exchange visit at Yale University, New Haven CT, USA, with the Interactive Machines Group led by Marynel Vázquez. |
| 2024 | SEQUEL and POLITE published at ICRA 2024. |
| 2024 | Human-Centric Autonomous Systems With LLMs for User Command Reasoning received the Best Student Paper Award at the WACV 2024 LLVM-AD Workshop. |
| 2024 | PREDILECT published at HRI 2024, using zero-shot language-model reasoning for preference-based RL. ACM page. |
| 2023 | VARIQuery presented at IROS 2023 in Detroit, Michigan. |
| 2023 | Aligning Human Preferences with Baseline Objectives in Reinforcement Learning presented at ICRA 2023 in London, United Kingdom. |
| 2021 | Human-feedback shield synthesis for perceived safety in deep reinforcement learning published in IEEE Robotics and Automation Letters. |
We introduce Velocity-Field Disagreement (VFD) for epistemic uncertainty in flow-based VLAs and SAVE, an uncertainty-guided active fine-tuning framework. On LIBERO, SAVE reduces expert data needs by at least 22% while improving failure awareness and adaptation.
Accepted to the RSS 2026 Diff4RL Workshop; arXiv:2606.18043, 2026
Introduces Velocity-Field Disagreement (VFD) for epistemic uncertainty in flow-based VLAs and SAVE, an uncertainty-guided active fine-tuning framework that needs at least 22% fewer expert demonstrations than baselines.
In IEEE International Conference on Robotics and Automation (ICRA), 2026
Reframes preference-based RL as multi-objective learning: language explanations are parsed into objective-specific labels, weights, and highlights for scalarized policy optimization.
In International Conference on Machine Learning (ICML), 2026; also accepted to the ICLR 2026 Workshop on Scaling Post-training for LLMs (SPOT)
Introduces Self-Distillation Policy Optimization (SDPO), converting rich environment feedback into dense self-distillation signals for more sample-efficient RL with language models.
In IEEE International Conference on Robotics and Automation (ICRA), 2025
Adapts reward functions with low-rank style parameters to support preference-based robot adaptation from limited feedback while mitigating catastrophic reward forgetting.
In IEEE International Conference on Robotics and Automation, 2024
Improves preference-learning sample efficiency by augmenting human feedback with synthesized preference queries from latent interpolation.