Active Mapping 3D

Autonomous 3D reconstruction using live incremental Gaussian Splatting on a PX4 drone in Gazebo simulation. The drone flies orbit patterns around objects, captures RGB-D images, and builds 3D Gaussian Splat reconstructions in real-time using gsplat.

Dependencies

• ROS 2 Jazzy
• PX4-Autopilot (with custom airframe and models — see simulation/README.md)
• Micro-XRCE-DDS-Agent
• Python 3.12 venv with:
  • torch >= 2.6 (CUDA 12.4)
  • gsplat == 1.5.3
  • opencv-python
  • matplotlib
  • numpy < 2

Clone

git clone --recursive https://github.com/DREAMS-lab/active_mapping_3D.git

Build

cd active_mapping_3D
source /opt/ros/jazzy/setup.bash
colcon build --packages-select mapping gsplat --symlink-install

Simulation Setup

Before the first run, copy custom PX4 models, airframe, and world files into your PX4-Autopilot installation. See simulation/README.md for instructions.

Run

Three terminals:

T1 — PX4 SITL:

cd ~/PX4-Autopilot
PX4_GZ_NO_FOLLOW=1 PX4_GZ_WORLD=sample_15016 make px4_sitl gz_px4_gsplat

T2 — DDS Agent:

MicroXRCEAgent udp4 -p 8888

T3 — Active Reconstruction:

source active_mapping_3D/install/setup.bash
ros2 launch mapping active_recon.launch.py rock_x:=0.0 rock_y:=8.0

This will:

1. Take off and fly two orbit passes around the rock (24 waypoints each)
2. Train a 3D Gaussian Splat model live during flight
3. Open an OpenCV splat viewer (drag to orbit, scroll to zoom)
4. Show voxel coverage and camera poses in RViz2
5. Save evaluation outputs (renders, PSNR/SSIM, plots) after landing

Launch Parameters

Parameter	Default	Description
rock_x	0.0	Rock position X (NED, meters)
rock_y	8.0	Rock position Y (NED, meters)
orbit_waypoints	24	Waypoints per orbit pass
orbit_radius	2.0	Orbit radius (meters)
iters_per_keyframe	500	Training iterations per round
window_size	10	Spatial window size for training
max_gaussians	200000	Maximum gaussian count
pts_per_frame	40000	Points sampled per keyframe
enable_viewer	true	OpenCV splat viewer
enable_orbslam3	false	ORB-SLAM3 (requires separate build)

Output

Data is saved to data/mapping/run_NNN/:

• images/ — RGB frames
• depth/ — Depth frames (uint16, millimeters)
• transforms.json — Camera poses and intrinsics
• model/ — Gaussian splat checkpoint
• eval/ — Rendered views, loss curves, PSNR/SSIM plots, coverage plots

Architecture

Single-process, three-thread design:

• Thread 1 (ROS main): Flight FSM, keyframe capture, PX4 offboard control, voxel coverage, RViz2
• Thread 2 (CUDA optimizer): Incremental 3DGS training with gsplat, spatial windowed training, adaptive iteration budgeting
• Thread 3 (Viewer): OpenCV splat viewer with mouse orbit controls

Camera: 640x480 @ 30fps, fx=fy=465.7412