[go: up one dir, main page]

Roham Z. Nobari

SLAM 3D vision egocentric perception

  • Computer Vision Team Lead, microagi
  • Graduate Student Researcher, ETH Zurich CVG
  • M.Sc. AI, UZH and special student (CS) at ETH Zurich

Zurich, Switzerland

rohamzn.com · rzendehdelethzch · github.com/rzninvo · linkedin.com/in/rohamzn

Animated SLAM run: a camera maps landmarks while its pose estimate drifts, then a loop closure is detected and pose-graph optimization pulls the map back into alignment. Animated structure-from-motion turntable: a point map of a small temple turning inside a ring of eight keyframe camera frustums.

About

Portrait of Roham Zendehdel Nobari

I work on SLAM (simultaneous localization and mapping) for egocentric, first-person video, where hands, people and objects rarely stay still, and on turning that video into 3D data that robots can learn from. At ETH Zurich's Computer Vision and Geometry Group (Prof. Marc Pollefeys), Mahbod Tajdini and I co-designed and built microSLAM, dense monocular SLAM for dynamic egocentric scenes, as a two-person project supervised by Daniel Barath. It ranks 1st among monocular-only entries on the LaMAria leaderboard and runs in production at microagi (announcement), where I lead the team that turns its output into robot-learning data.

Before microSLAM, I worked on localization and 3D mapping: I co-developed LangLoc (ECCV 2026), which finds where you are in a known indoor scene from a description of what you see, and co-built SpotMap, an updatable 3D scene graph for the Spot robot. I'm doing all this alongside an M.Sc. in AI at the University of Zurich, with computer science courses at ETH Zurich as a special student.

Education

News

  1. Sep 2026

    Shayan Sooratgar and I received the Touché 2026 Award at CLEF 2026 in Jena for the highest event extraction effectiveness in the Causality Extraction task, where we presented our paper Shrome (paper, slides).

  2. Sep 2026

    microagi announced microSLAM, our monocular SLAM for egocentric video (announcement).

  3. Aug 2026

    Became Computer Vision Team Lead at microagi.

  4. Jun 2026

    LangLoc accepted to ECCV 2026 as a main-conference poster, co-first author (arXiv, code).

  5. Apr 2026

    Joined microagi as Perception Engineer (SLAM, contract).

  6. Mar 2026

    Became Graduate Student Researcher at ETH Zurich's CVG, after my LangLoc master project there.

  7. Nov 2025

    Joined the ETH Robotics Club's Humanoid Division, the team behind XALUN, as Robot Learning Engineer.

Publications

  1. LangLoc overview: a description of what you see goes through scene retrieval, fine localization and yes/no dialogue to the final pose.

    LangLoc: “Tell Me What You See”

    S. K. Panwar*, R. Zendehdel Nobari*, S. F. Y. Lau*, A. B. R. Shaik*, M. Günther, M. Pollefeys, D. Barath (* equal contribution)

    ECCV 2026 co-first author main conference, poster

    LangLoc finds where you are in a known building from a description of what you see, with no camera: it matches your words to a 3D scene graph of the rooms, estimates your position and viewing direction, and asks yes/no questions to narrow it down.

  2. Shrome at Touché: Soft-Vote Ensembling and Counter-Causal Augmentation for Causality Extraction

    R. Zendehdel Nobari*, S. Sooratgar* (* equal contribution)

    CLEF 2026 Working Notes (Touché shared task, CEUR-WS), to appear · talk in Jena, Sep 2026

    Touché 2026 Award at CLEF 2026: highest event extraction effectiveness in the Causality Extraction task

    The task: given news text, decide whether it makes a causal claim, extract the two events involved, and classify whether one is said to cause the other, not to cause it, or neither.

    Our system combines several models by soft voting and adds counter-causal training sentences (ones that deny a causal link).

Experience

microagi · egocentric data for robot learning

  1. Computer Vision Team Lead

    Aug 2026 – Present

    • Co-own microSLAM in production with M. Tajdini: the monocular system we built at CVG (below) and its visual-inertial variant.
    • Lead the team turning microSLAM output into robot-learning data: Gaussian splatting, meshes for Isaac Sim, 3D object tracking, 6-DoF object pose estimation, 3D scene flow.

  2. Member of Technical Staff (part-time)

    Apr 2026 – Aug 2026

    • Co-built microSLAM's tightly coupled visual-inertial variant with M. Tajdini: camera and IMU measurements are optimized jointly (Ceres), with loop closure to remove drift and GPU-accelerated solves (cuDSS).

  3. Perception Engineer (SLAM), contract

    Apr 2026

ETH Zurich, Computer Vision and Geometry Group (CVG) · supervisors: Prof. Marc Pollefeys, Dr. Daniel Barath

  1. Graduate Student Researcher (part-time)

    Mar 2026 – Present

    • Co-designed and built microSLAM with Mahbod Tajdini: dense monocular SLAM that stays accurate when hands, people and objects move, by tracking moving objects separately and mapping each one as its own instance; targeting CVPR 2027.
    • We flagged inflated LaMAria scores, our own included: a bug in the scorer's trajectory alignment let less accurate trajectories score higher. The maintainers fixed it and re-scored every affected entry (PR #22).
    • On the re-scored LaMAria leaderboard, microSLAM ranks 1st among monocular-only (no-IMU) entries on all five challenges; three rival monocular entries there are our own reruns.

  2. Master Project Student

    Jun 2025 – Mar 2026

    • Co-developed LangLoc (ECCV 2026): you describe what you see, and it finds the room, your position and your viewing direction in a known building, asking yes/no questions when the description is ambiguous.
    • Built its dataset of 13,000+ pose-indexed descriptions of indoor scenes, written by an LLM from a diverse set of camera views.

ETH Robotics Club, Humanoid Division · team behind XALUN, Switzerland's first student-built humanoid

  1. Robot Learning Engineer (part-time)

    Nov 2025 – Present

    • Set up whole-body teleoperation of a Unitree G1 humanoid through a VR headset (NVIDIA GEAR-SONIC, PICO) with the team, and collected demonstrations to fine-tune NVIDIA's Isaac GR00T robot foundation model.
    • Added live camera streaming (ZED Mini and wrist cameras, Rerun) and fixed a buffer setting that had halved the wrist camera's frame rate.
    • Co-reconstructed the club's hangar as a 3D mesh for simulation (real-to-sim); a Gaussian-splatting map is in progress.

ETH Zurich, Ecosystems and Landscape Evolution Group (D-USYS) · supervisors: Prof. Loïc Pellissier, Dr. Camille Albouy

  1. Research Assistant (part-time)

    Apr 2025 – Jun 2026

    • Extended the lab's scripts into SeeDNAP, an open-source pipeline that turns raw environmental-DNA (eDNA) sequencing data into biodiversity records.

Projects

Teaching & Service

Awards

Skills

3D vision & SLAM
visual and visual-inertial SLAM, bundle adjustment, pose-graph optimization, visual localization, 3D reconstruction, 3D scene graphs
Machine learning
deep learning, graph neural networks, vision foundation models, imitation learning, reinforcement learning
Robotics
ROS 2, sim-to-real, simulation (Habitat-Sim, Flightmare, Autoware/AWSIM)
Robot platforms
Unitree G1 humanoid, Boston Dynamics Spot, SO-101 arm, quadrotor
Code & tools
Python, C++, CUDA, PyTorch, Ceres, OpenCV, Open3D, Docker, Linux, Git
Languages
English (C2), German (B1), Persian (native), Japanese (elementary)