ETH D-INFK · Computer Vision and Geometry Group  ·  Graduate Student Researcher (Mar 2026 – Present)

A state-of-the-art Dynamic SLAM system developed at CVG with Dr. Dániel Béla Baráth, advancing real-time localization and mapping in scenes with moving objects, where classical SLAM assumptions break down.

Visual SLAM 3D Vision Dynamic Scenes C++ · Python

ETH CVG Lab × UZH AI/ML Group

The first pipeline for fine-grained indoor localization from natural language alone, without any camera input. A dual-branch GATv2 + CLIP architecture for scene retrieval (+8 pp Top-1 over SOTA), a visibility-based floor-grid scoring module that reaches ~1 m median error, and a Bayesian dialog module that drives median error down to 7 cm via targeted yes/no questions.

PyTorch CLIP GNNs Bayesian Inference

ETH Ecosystem & Landscape Evolution Lab

A Transformer-based mask modeling framework (inspired by Pl@ntBERT) on phylogenetic embeddings to detect ecologically suitable but absent marine species, quantifying anthropogenic impact at scale. Also built an end-to-end FASTQ → DADA2 → GBIF pipeline in collaboration with marine ecologists.

Transformers Bioinformatics eDNA R · Python

ETH Robotics Club

Implementing TWIST in NVIDIA IsaacLab to train a Unitree G1 locomotion policy on AMASS / MoCap data with robust sim-to-real transfer. Fine-tuning NVIDIA GR00T N1.6 on teleoperation data and deploying SONIC for real-time whole-body control, alongside a low-cost VR-based teleoperation rig built with PICO headsets.

IsaacLab RL Sim2Real Unitree G1

ETH CVG Lab

Onboard 3D perception for the Spot robot: dense RGB-D SLAM with TSDF volumetric fusion from a single monocular camera, plus a dynamic scene graph generator powered by OpenMask3D for open-vocabulary instance segmentation, letting the robot semantically understand and update its world during interaction.

SLAM TSDF OpenMask3D Scene Graphs

Microsoft Spatial AI Lab

A voice-based navigation assistant for smart glasses that guides users with context-aware, landmark-based verbal cues instead of metric instructions. Combines image-based localization, semantic landmark extraction, and mesh-aligned grounding inside Habitat-Sim, evaluated on a 3D mesh of the ETH HG building.

Habitat-Sim VLMs Smart Glasses Spatial AI

UZH Robotics & Perception Group

A PPO-based vision policy for an autonomous "camera drone", trained in Flightmare / Agilicious to track dynamic targets while keeping safe flight dynamics. Deployed on real hardware through a 60 Hz ROS control loop, with extensive domain randomization for robust sim-to-real transfer.

PPO Flightmare ROS Sim2Real