Guilhem Carmouze

UPSSITECH, engineering school of the University of Toulouse. French engineering degree, Master’s level. Fifth and final year in 2026 to 2027.

UPSSITECH, University of Toulouse

Artificial Intelligence Research Center (AIRC), AIST, Tsukuba, Japan. 29-page report in English: “Creation of a 360° Navigation Dataset Using 3D Gaussian Splatting”.

• Built ArtiFixer-360, a pipeline from a plain pinhole video to a 360° video: COLMAP, a 3DGRUT Gaussian scene, a world-locked rig of 14 overlapping 110° views repaired jointly by a 14B video diffusion model, then distilled back into the scene, with reference-free quality gates. +23,602 lines and 119 tests over NVIDIA’s upstream ArtiFixer; runs on 4-GPU nodes of the ABCI cluster.
• Depth-aware synchronisation lowered the cross-view depth-overlap error of the repaired views by 27% (0.0340 to 0.0247) before distillation. On a 117-frame reference run the temporal warp error (MAE, raw renders against output) fell from 0.037 to 0.020. The full 154-frame run failed my own acceptance gates and led to a geometry-first redesign.
• Earlier phases: A* on a 5 cm occupancy grid and 4096 × 2048 panoramas rendered from six pinhole views, logged with position and heading, in a DISCOVERSE and MuJoCo simulation; a ROS 2 Humble interface prepared for a visual navigation model on the Kachaka robot (on main the model is not wired in; no run on the real robot is claimed).
• Also worked on SVLR (training-free visual language robotics, CNRS-AIST JRL), a modular pipeline that turns an instruction into robot actions without training.

AI tools were used for research, code and spell-checking, as declared in the report; not for running experiments or producing results.

Codeartifixer-360-pipelinenav_3dgs_panoKachakaNavigation

Industry 4.0 smart factory (Usine 4.0), with my class at UPSSITECH.

• My part: the simulation and tooling layer. A 2D Pygame simulator with a fit-to-track camera, a typed CSV cone-track loader and a configurable field-of-view sensor model that selects the visible cones.
• Also worked on cone-detection models (PyTorch) and on image-processing and control modules (ROS, Python, C++). The planners are my teammates’ work.

CodeTLSe_Racing_DriverlessPathPlanning

Team of six: Arduino motor control, Raspberry Pi camera, LiDAR mapping with ICP, voice commands, ball tracking.

• My part: the web interface, a single-page app that drives the robot over the Web Bluetooth API; the Raspberry Pi MJPEG camera stream; the algorithm that turns the ball’s image coordinates into drive commands to keep it centred; the timing of split voice commands.
• The semester before, with a classmate: a colour-ball detector in pure C11, without OpenCV.

CodePFR2 (team repository)PFR