• RP2040 (Raspberry Pi Pico) microcontroller publishing at 10 Hz over GPIO and ROS.
• LiDAR and camera are now solidary (rigidly mounted together).
• All ROS2 source repos have been merged.
• Extrinsic calibration algorithm selected.
• ZFS mount set up with deduplication, compression, and automatic daily snapshots for data recording.
• Installed and trained for FPV flying — not directly used in the project but necessary to safely operate the drone.
• ESC integration with rover and ROS2 completed.

• Updated the dev log.
• Finished comparing Point Cloud Data (PCD) from the LiDAR against Gaussian Splatting reconstructions.
• Got a proper version of georefereced Gaussian splats working using Cesium.
• The RTK GPS is working and publishing correctly on ros.
• Micro ros is functionnal
• New updated order list (THe motors are not rated for 6S battries and some other )

3D Reconstruction pipeline experiments:

• GLomap works significantly better and faster than Colmap. Key advantage: optimises over all images simultaneously rather than incrementally. Limitation: struggles when images lack sufficient overlap or parallax.
• Lichtfeld outperforms Nerfstudio for this use case. Tested various parameters — GSP+ gives the best results but behaves quite differently from others.
• Lichtfeld trained with camera poses estimated from LiDAR + odometry (via Colmap) works well and is faster than a full GLomap run. In this setup the LiDAR is used for positioning only, not as a point cloud reference.
• Next step: proper camera–LiDAR extrinsic calibration, then use LiDAR data either combined with Depth Anything v3 or standalone as a geometric prior to reduce free-floating noise.

PCD vs Gaussian Splatting comparison:

Wrote a script to compare LiDAR point clouds against GS reconstructions. Some similarity is visible but the GS output is still too noisy for reliable filtering.

On-the-fly NVS (graphdeco-inria/on-the-fly-nvs):

Among the best viewer results so far, but not yet fully exploitable in practice.

Gerboise (rover) — first drive:

Wheels, frame and VESC ESC are working. ROS2 drivers and Xbox controller confirmed operational.

• Tricity Labs page created on the site.
• Rover project scoped: GPS-RTK module selected, chassis design started.
• Drone project scoped: camera-based localisation pipeline research started.
• Bought an Ardusimple RTK-GPS
• Had a LiDAR laying around (Livox Avia)

20/04/2026

• Part list for drone, camera and rover finalised.
• RTK module received — experiments to start soon.

21/04/2026 — Main hardware order placed (AliExpress, ~676 €)

Additional hardware (separate purchases):

• Ardusimple RTK-GPS — 270 €
• VESC Flipsky ESC (rover) — 180 €
• Industrial camera MV-CU013-AOGC + 6mm lens (hardware sync) — 200 € + 32€ Taxes
• Ammunition box (battery enclosure) — 15 €
• 18650 cells for drone and rover — ~30 € (considering: LiFePo4 ×14 1100mAh/30A, Li-ion ×10 3200mAh/10A, Li-ion ×5 2100mAh/30A)

• Initial research on offroad autonomous navigation approaches (ROS2, Nav2, custom planners).
• Benchmarked GPS-RTK modules for cost vs. accuracy trade-offs.
• Evaluated open-source photogrammetry pipelines (OpenDroneMap, Colmap) for 3D reconstruction.
• Studied GPS-denied localisation methods: ORB-SLAM3, OpenVINS.