How hzai-zjnu/duag-slam works
DUAG-C runs independent RGB-D Gaussian-SLAM per robot, then uses inter-robot loop closures to align poses and fuse matched map elements with Fisher-information weighting. The main runtime entry is the experiment runner; each RobotNode owns frame processing, message handling, and consensus dispatch. The README also documents dataset loading, uncertainty estimation, real-time visualization, and result analysis. Some of those branches are represented at their documented directory or module level because their implementation was not excerpted.
Main entry points: robot_node.py Robot orchestration · run_duag_c.py Experiment runner · local_slam_wrapper.py Local SLAM · renderer.py Gaussian renderer
Components 19
Experiment runtime
- Experiment runner
experiments/run_duag_c.py - Dataset loading
experiments/data_loaders.py - Robot orchestration
core/pipeline/robot_node.py - Map and pose types
core/types.py
Local mapping
- Local SLAM
core/pipeline/local_slam_wrapper.py - Gaussian map model
extracted/gs_slam/gaussian_model.py - Gaussian renderer
extracted/gs_slam/renderer.py - CUDA rasterizer
extracted/gs_slam/rasterizer
Decentralized consensus
- Loop closure matching
core/consensus/matching.py - Riemannian ADMM
core/consensus/riemannian_admm.py - SE3 Lie algebra
core/consensus/lie_algebra.py - Gaussian map fusion
core/consensus/gaussian_consensus.py - Consensus convergence
core/consensus/convergence.py
Uncertainty and communication
- FIM estimation
core/uncertainty/hessian_estimator.py - Uncertainty propagation
core/uncertainty/propagation.py - Robot communication
extracted/swarm_comm
Visualization and results
- Live visualization
core/visualization/realtime_viewer.py - Result analysis
experiments/analyze_results.py
Other components
- Experiment operator
Connections 23
- Experiment operator → Experiment runner: runs
README.md - Experiment runner → Robot orchestration: creates agents
experiments/run_duag_c.py - Experiment runner → Local SLAM: configures
experiments/run_duag_c.py - Experiment runner → Live visualization: enables
experiments/run_duag_c.py - Experiment runner → Dataset loading: loads sequences
README.md - Robot orchestration → Local SLAM: processes frames
core/pipeline/robot_node.py - Robot orchestration → Robot communication: receives messages
core/pipeline/robot_node.py - Robot orchestration → Loop closure matching: dispatches matching
core/pipeline/robot_node.py - Robot orchestration → Riemannian ADMM: updates poses
core/pipeline/robot_node.py - Robot orchestration → Gaussian map fusion: fuses maps
core/pipeline/robot_node.py - Robot orchestration → Consensus convergence: checks convergence
core/pipeline/robot_node.py - Local SLAM → Gaussian map model: updates map
core/pipeline/local_slam_wrapper.py - Local SLAM → Gaussian renderer: renders views
core/pipeline/local_slam_wrapper.py - Gaussian renderer → Gaussian map model: reads Gaussians
extracted/gs_slam/renderer.py - Gaussian renderer → CUDA rasterizer: rasterizes
extracted/gs_slam/renderer.py - Local SLAM → FIM estimation: estimates uncertainty
README.md - Gaussian map fusion → Loop closure matching: matches Gaussians
core/consensus/gaussian_consensus.py - Gaussian map fusion → Uncertainty propagation: weights fusion
core/consensus/gaussian_consensus.py - Riemannian ADMM → SE3 Lie algebra: uses SE3 operations
core/consensus/riemannian_admm.py - Riemannian ADMM → Uncertainty propagation: uses pose uncertainty
core/consensus/riemannian_admm.py - Uncertainty propagation → SE3 Lie algebra: uses transforms
core/uncertainty/propagation.py - Robot orchestration → Map and pose types: stores map state
core/pipeline/robot_node.py - Experiment runner → Result analysis: analyzes outputs
README.md