How rpng/mins works
Project Type and Purpose
• MINS is a ROS-based, C++ multisensor fusion system for inertial navigation.
• It ingests IMU, wheel odometry, camera, GNSS, LiDAR (and optionally VICON or loop-closure) data, runs a tightly-coupled filter to estimate pose, and supports simulation and offline evaluation.
High-Level File Structure
• mins/: core catkin package
• config/ – YAML configs for datasets and sensors
• launch/ – ROS launch files (simulation, rosbag replay, subscription)
• src/ – modules broken into:
· core (ROS wrappers & SystemManager)
· init (sensor initializers)
· options (parsing per-sensor options)
· sim (Simulator & visualizer)
· state (state propagation & representation)
· update (sensor-specific update plugins)
· utils (logging, timing, helper routines)
· run_*.cpp (entry points: run_bag, run_simulation, run_subscribe)
• mins_data/: sample datasets, ground truths, floorplans
• mins_eval/: offline evaluation tools (consistency, accuracy, timing)
• thirdparty/: bundled libraries (OpenVINS core & eval, libnabo, libpointmatcher/ikd-tree)
• Top-level: CMakeLists.txt, Dockerfiles, GitHub CI workflow
Readme-Inferred Architecture
• At runtime MINS launches one of three ROS nodes:
a) Simulation node (Simulator → publishes synthetic sensor topics)
b) ROS-bag replay node (reads bag → republishes topics)
c) Subscription node (subscribes existing topics)
• All feed into a central SystemManager via ROSSubscriber.
• SystemManager drives a filter that alternates:
• Propagator (integrates IMU on manifold)
• Updaters (camera, GNSS, LiDAR, wheel, VICON plugins)
• State estimates are published via ROSPublisher.
• Evaluation modules (mins_eval) ingest both published estimate and ground truth for offline plots.
Main Components
a. Data Sources
– Real sensors (IMU, wheel, camera, GNSS, LiDAR, VICON) via ROS topics
– Simulation engine (Simulator in mins/src/sim)
– ROS-bag replay
b. Middleware
– ROS 1 master & topic bus
– ROSSubscriber / ROSPublisher (mins/src/core)
c. Core Estimator
– SystemManager (mins/src/core)
– State Propagator (mins/src/state)
– Sensor Updaters (mins/src/update/{imu,cam,gps,lidar,wheel,vicon})
– Initializer modules (mins/src/init)
– Options/config loader (mins/src/options)
d. Utilities
– Logging & timing (mins/src/utils)
e. Offline Evaluation
– mins_eval package: alignment, error plots, timing comparisons
f. Third-Party Libraries
– OpenVINS core & eval
– libnabo / ikd-Tree / libpointmatcher for LiDAR ICP
Relationships and Data Flow
• Within SystemManager: loop of Propagate → Update(plugin) → Publish
• Options feed into both Initializer and SystemManager at startup
• Simulation node sits parallel to real data sources, publishing onto same topics
• Evaluation tools consume topics/log files offline
Architectural Patterns & Principles
• ROS node/component-based modularity
• Plugin style for each sensor updater
• High-order state interpolation on manifold
• Dynamic cloning of states for multiple sensors
• Separation of real vs. simulated data paths
Technologies & Dependencies
• ROS 1 (Melodic/Noetic), catkin, CMake
• C++17, Eigen, YAML configs
• Docker for CI images
• Third-party libs (OpenVINS, ikd-Tree, libpointmatcher)
Essential Elements to Capture
– Clear component labels (with source directories)
– Arrowed data flows (sensor → estimator → output → evaluation)
– Startup/config dependencies (Options, Initializer)
– Distinct runtime nodes (run_simulation, run_bag, run_subscribe)