Portrait of Simon Roy

Simon Roy

I work on robot learning: world models, language-conditioned planning and vision-language reward models.

I am a PhD student at Polytechnique Montréal and Mila, advised by Giovanni Beltrame. I am also co-founder and CTO of Aquilix, where we put computer vision to work on construction and industrial sites.

News

  1. Jul 2026 New first-author preprint on searching satellite image archives with text queries.
  2. Jun 2026 Two preprints with collaborators: LAGO, on latent goal prediction for model-based planning, and VISTA, on scale-aware visual navigation.
  3. May 2026 Started my PhD in robot learning at Polytechnique Montréal and Mila.
  4. Dec 2025 First-author preprint revisiting the learning objectives of vision-language reward models.
  5. Sep 2025 Teaching assistant for INF3995, the third-year robotics design project at Polytechnique.
  6. Jul 2025 Co-founded Aquilix, where I am CTO.

Company

Co-founder and CTO, since July 2025

Aquilix builds self-contained AI cameras that detect risks and critical events in real time on construction sites, rail yards and remote sites. A unit installs in a few hours and runs over LTE, with no site internet needed.

  • Lead the technology end to end: camera systems, on-device vision models and the analytics platform.
  • Turn research in computer vision and vision-language models into systems that run unattended in the field.

Visit aquilix.ca

Research

  1. Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently

    Simon Roy, Mark Bong, Giovanni Beltrame

    arXiv preprint, Jul 2026 Abstract PDF
    BibTeX
    @misc{roy2026finding,
      title         = {Finding Change in Satellite Archives from Text: How to Combine Before-and-After Images Efficiently},
      author        = {Simon Roy and Mark Bong and Giovanni Beltrame},
      year          = {2026},
      eprint        = {2607.28571},
      archivePrefix = {arXiv},
      url           = {https://arxiv.org/abs/2607.28571},
    }
  2. Latent Goal Prediction from Language for Model-Based Planning

    Samuel Barbeau, Simon Roy, Giovanni Beltrame, Christian Desrosiers, Nicolas Thome

    arXiv preprint, Jun 2026 Abstract PDF
    BibTeX
    @misc{barbeau2026latent,
      title         = {Latent Goal Prediction from Language for Model-Based Planning},
      author        = {Samuel Barbeau and Simon Roy and Giovanni Beltrame and Christian Desrosiers and Nicolas Thome},
      year          = {2026},
      eprint        = {2606.20627},
      archivePrefix = {arXiv},
      url           = {https://arxiv.org/abs/2606.20627},
    }
  3. VISTA: Scale-Aware Visual Navigation via Action History Conditioning

    Maeva Guerrier, Koki Kobayashi, Simon Roy, Jana Pavlasek, Giovanni Beltrame

    arXiv preprint, Jun 2026 Abstract PDF
    BibTeX
    @misc{guerrier2026vista,
      title         = {VISTA: Scale-Aware Visual Navigation via Action History Conditioning},
      author        = {Maeva Guerrier and Koki Kobayashi and Simon Roy and Jana Pavlasek and Giovanni Beltrame},
      year          = {2026},
      eprint        = {2606.17294},
      archivePrefix = {arXiv},
      url           = {https://arxiv.org/abs/2606.17294},
    }
  4. To Select or not to Select, that is the Question: Distilling Robot Skill Prediction into a Small Ensemble

    Haechan Mark Bong, Simon Roy, Euhid Aman, Giovanni Beltrame

    arXiv preprint, May 2026 Abstract PDF
    BibTeX
    @misc{bong2026select,
      title         = {To Select or not to Select, that is the Question: Distilling Robot Skill Prediction into a Small Ensemble},
      author        = {Haechan Mark Bong and Simon Roy and Euhid Aman and Giovanni Beltrame},
      year          = {2026},
      eprint        = {2605.21242},
      archivePrefix = {arXiv},
      url           = {https://arxiv.org/abs/2605.21242},
    }
  5. Revisiting the Learning Objectives of Vision-Language Reward Models

    Simon Roy, Samuel Barbeau, Giovanni Beltrame, Christian Desrosiers, Nicolas Thome

    arXiv preprint, Dec 2025 Abstract PDF
    BibTeX
    @misc{roy2025revisiting,
      title         = {Revisiting the Learning Objectives of Vision-Language Reward Models},
      author        = {Simon Roy and Samuel Barbeau and Giovanni Beltrame and Christian Desrosiers and Nicolas Thome},
      year          = {2025},
      eprint        = {2512.20675},
      archivePrefix = {arXiv},
      url           = {https://arxiv.org/abs/2512.20675},
    }

Full list on Google Scholar

Projects

  • Goat Brain interface with the chat and the generated behavior tree

    Goat Brain

    A speech-to-action agent for a mobile robot: a LoRA-tuned LLM with retrieval turns spoken instructions into ROS 2 behavior trees, with a visual interface to follow what the robot is doing.

    CodeVideo

  • 3D semantic map of an indoor scene with detected objects in red

    Shepherd

    Open-vocabulary semantic mapping for a mobile robot. Segments and captions objects with FastSAM, CLIP and BLIP, then answers natural-language queries over the map.

    Code

  • Tracked rescue robot with a manipulator arm beside a rescue dummy

    Search-and-rescue robot autonomy

    Club Capra, ÉTS

    Autonomy stack for a tracked search-and-rescue robot: ROS 2 architecture in Docker, 3D SLAM, exploration, object tracking, arm inverse kinematics and Gazebo simulation. Best in class for search and rescue at EnRicH 2023; competed at ELROB 2024.

    Club Capra

  • Game level with a quantum circuit puzzle below it

    Entangled Explorer

    QuantumÉTS

    A 2D puzzle platformer that teaches quantum computing concepts to a general audience.

    PlayVideo

All projects, including side projects

Experience

  1. Computer vision and machine learning engineer

    NIO Robotics

    Developed computer vision and machine learning methods to identify the paintable areas of a wall for an autonomous wall-painting robot.

  2. Quantum machine learning research

    Université de Sherbrooke

    Developed fermionic machine learning algorithms and data pipelines in the quantum machine learning group.

  3. Autonomy software lead

    Club Capra, ÉTS

    Led the autonomy stack of a search-and-rescue robot: containerized ROS 2 architecture, SLAM and EKF-based navigation, simulation, and a vision-language pipeline for 3D object mapping. Competed at EnRicH 2023 and ELROB 2024.

  4. Robotics software engineer

    Canadian Space Agency

    Built mission-planning tools for the Canadian lunar rover (Django, React), worked on path planning (RRT*, genetic algorithms) and operated the lunar map server and its CI/CD.

  5. Co-founder and president

    QuantumÉTS

    Founded the ÉTS quantum computing club: about 20 active members, workshops for more than 200 students, and hackathon teams ranked in the top 5% at QHack two years in a row.

  6. Software developer

    Orolia (now Safran)

    Signal processing and generation features for the Skydel GNSS simulator, in C++, Python and MATLAB.

  7. Hardware engineering

    Lion Électrique

Teaching

  1. Teaching assistant, INF3995

    Polytechnique Montréal

    Third-year team design project in robotics. Taught ROS 2, Docker and Linux, introduced course rules on responsible use of AI tools, and ran a rate-limited inference server so every team had fair access to shared models.

  2. Teaching assistant, MTI882B

    École de technologie supérieure

    Algorithmic aspects of quantum computing. Wrote the course notebooks and taught PennyLane.

Education

  1. PhD, Robot learning

    Polytechnique Montréal and Mila GPA 4.0/4.0

    Advised by Giovanni Beltrame. World models, language-conditioned planning and vision-language reward models.

  2. M.Sc., Artificial intelligence

    Polytechnique Montréal and Mila GPA 4.0/4.0

  3. B.Eng., Software engineering

    École de technologie supérieure (ÉTS) Graduated with honours

Distinctions

2024 Honorable mention from NVIDIA, ETH Zurich Quantum Hackathon
2024 Competed with Club Capra, ELROB, European Land Robot Trial
2023 Best in class, search and rescue, EnRicH, European Robotics Hackathon
2023–24 Top 5% of teams, two years running, Xanadu QHack
2022 3rd place, BIG Quantum Hackathon
2020 1st place, Halifax, NASA Space Apps Challenge

Contact

Email is the best way to reach me about research, teaching or Aquilix.