Repository navigation
v1.0.0
🚀 DDDMR Navigation Stack v1.0.0 — Official First Release
Celebrating 5 Years of Research, Engineering, and Real-World Field Testing
We are thrilled to announce the first official release of the DDDMR Navigation Stack (v1.0.0)!
Born out of five years of continuous development, simulation, and extensive field testing across diverse mobile platforms, DDDMR (3D Mobile Robot Navigation) delivers a complete, end-to-end autonomous navigation solution specifically engineered for complex 3D environments.
🌟 Why DDDMR Navigation?
Traditional 2D navigation stacks (such as ROS 2 Nav2 standard setups) excel on flat surfaces, but struggle in multi-dimensional real-world environments: multi-level structures, steep factory ramps, wheelchair-accessible inclines, uneven terrains, and complex multi-layer obstacles.
DDDMR bridges this gap. It brings the familiar, battle-tested workflow of 2D mobile robot navigation into true 3D space:
- 3D SLAM & Mapping: Map multi-layer environments and edit pose graphs interactively.
- Pose-Graph 3D Localization: Relocalize reliably in vast areas without prohibitive memory or compute costs.
- True 3D Perception: Detect, mark, track, and clear 3D obstacles with multi-sensor support.
- 3D Global & Local Planning: Plan paths across multi-level structures and ramps while respecting real robot kinodynamic and motor limits.
- Cost-Effective Hardware Compatibility: Engineered from the ground up to run smoothly on affordable hardware—such as 16-line LiDARs, consumer IMUs, and edge computing boards (e.g., Intel NUC and NVIDIA Jetson Orin Nano).
📦 What's Included in This Release
🗺️ 1. 3D SLAM & Pose Graph Pipeline (dddmr_lego_loam)
- ROS 2 Humble Native: Full rewrite and optimization of LeGO-LOAM-BOR for ground vehicles, supporting both online SLAM and offline rosbag processing.
- Deep LeGO-LOAM (YOLO11 + TensorRT): Integrated real-time deep learning dynamic obstacle detection to filter out humans and moving objects, preventing map ghosting and corruption.
- Any-Angle LiDAR Mounting: Native support for horizontal, tilted (e.g., RoboSense Airy pitched 45°), inverted/rolling (e.g., Livox Mid-360 rolled 180°), and complex roll-and-yaw configurations (e.g., Hesai JT128 with 45° roll decoupling).
- Interactive Pose Graph Editor & Merge Tool: Dedicated RViz2 GUI panels to inspect keyframes, adjust loop closures via optimized ICP, and stitch multiple independent submaps together.
- Odometry & LiDAR Desync Tolerance: Resilient timestamp drift handling to avoid mapping halts during bag playback or communication lags.
📍 2. Scalable 3D Localization (dddmr_mcl_3dl)
- Pose-Graph-Driven 3D MCL: Loads lightweight LeGO-LOAM pose graphs instead of massive monolithic point cloud files, drastically reducing memory footprint.
- Submap Architecture for Large-Scale Maps: Dynamically loads localized active submaps (
2 * lidar_range + 2 * search_radius), enabling seamless localization across 500m × 500m+ terrains directly on an NVIDIA Jetson Orin Nano. - Redesigned Cluster-Based Particle Rating: Incorporates Euclidean Cluster Extraction to fairly weigh distant sparse returns against dense near-field points, paired with surface normal tracking to eliminate "virtual slipping" along uniform walls.
- Motion-Triggered Updates: Particle filters update strictly based on translation/rotation thresholds, reducing idle CPU usage to near zero.
👁️ 3. True 3D Perception & Obstacle Clearing (dddmr_perception_3d)
- Multi-Sensor Ecosystem: Plug-and-play support for multi-layer spinning LiDARs (Velodyne, Ouster, Leishen C16), scanning/solid-state LiDARs (Livox Mid-360, Unitree L1/G4), and depth cameras (Intel RealSense D455/D435, Luxonis OAK).
- Ray-Cast Obstacle Marking & Clearing: High-performance 3D ray casting with zero self-blocking, empty-observation clearing (
observation_clear_), and KDTree Euclidean fallback ensuring complete removal of dynamic obstacles. - Layered Zone Management: Utilities to define static layers, speed limit zones, and no-go / keep-out zones directly via point clouds. Includes an interactive RViz2 Point Cloud Zone Editor and deletion tool.
🧭 4. 3D Global & Local Path Planning
- Graph-Based 3D Global Planner (
dddmr_global_planner):- Employs graph-based 3D A* directly on terrain ground point clouds.
- Dynamically merges static terrain with real-time perception obstacles.
- Automated ground boundary detection keeps planned paths safely centered and away from cliff edges, ramps, and drop-offs.
- Modular 3D Local Planner (
dddmr_local_planner):- Plugin-based architecture supporting custom Trajectory Generators, Critics, and Recovery Behaviors.
- True 3D collision checking using 3D bounding cuboids against dense point clouds.
- Motor Constraint-Aware Generation: Dynamically models physical motor RPM boundaries (e.g., differential-drive saturation) to guarantee all velocity commands are physically achievable.
- OpenMP parallel trajectory sampling and MPPI (
expertScoring) controller integration.
🕹️ 5. Unified Kinematics & Navigation FSM (dddmr_p2p_move_base)
- Robust Finite State Machine: High-level orchestrator handling navigation goals, retries, recovery behaviors, and execution monitoring.
- Broad Kinematics Support:
- Differential Drive
- Omnidirectional
- Ackermann Steering (with automated sequential goal execution and configurable tolerances)
- Explicit Actuator Configuration: Strict definition (
ActuatorType::MOTORvsActuatorType::STEERING) guaranteeing clean separation betweenTwistandAckermannDrivecommand pipelines.
📐 6. 3D Odometry & State Estimation (dddmr_odom_3d)
- Fuses 2D planar velocity with filtered 3D IMU Euler angles (roll, pitch, yaw) for reliable 3D odometry estimation.
- Adheres strictly to standard robotics spatial conventions (REP-103 / REP-105: X-forward, Y-left, Z-up).
- Extensively verified on quadruped platforms (such as DeepRobotics Lite3) across irregular ground.
🧠 7. AI Semantic Segmentation (dddmr_semantic_segmentation & dddmr_trt)
- Real-time RGB-D semantic point cloud colorization and classification powered by DDRNet and TensorRT.
- Achieves 15 FPS on Jetson Orin Nano and 19 FPS on Jetson Orin AGX.
- Supports semantic exclusion filters (e.g., masking traversable sidewalks vs. non-traversable boundaries).
🖥️ 8. Simulation, CI/CD, & Developer Experience
- Gazebo Quadruped Simulation (
dddmr_beginner_guide): Out-of-the-box Gazebo simulation environment featuring the Unitree Go2 quadruped robot navigating multi-level structures. - Automated CI/CD Test Suite (
CICD_setup): Data-driven regression testing that plays back verified rosbag datasets across multiple LiDAR geometries (Airy 45°, C16 0°, Mid360 180°, JT128 45° roll) and scores accuracy using automated ICP. - Turnkey Docker Environments (
dddmr_docker):- Ubuntu 22.04 CPU
- Ubuntu 22.04 CUDA 12.6 + TensorRT 10.7 + cuDNN 9 + PyTorch 2.8.0
- NVIDIA Jetson L4T JetPack 5.1.3 & 6.x (r36.4.0)
- Full Gazebo simulation image (
x64_gz)
- Global Accessibility: Complete beginner's guide translated into 12 languages (English, Arabic, Traditional Chinese, Simplified Chinese, Russian, German, Spanish, Korean, Portuguese, Turkish, Vietnamese, French, Japanese).
🤖 Supported & Verified Hardware
| Category | Tested Hardware / Configurations |
|---|---|
| LiDARs | RoboSense Airy (Dome, tilted 45°), Leishen C16, Hesai XT16, Hesai JT128 (45° roll), Livox Mid-360 (flat & 180° inverted roll), Unitree 4D LiDAR L1/G4 |
| Cameras | Intel RealSense D455 / D435, Luxonis OAK-D |
| Compute | Intel NUC (x86_64), NVIDIA Jetson Orin Nano (ARM64), NVIDIA Jetson Orin AGX |
| Platforms | Wheeled Mobile Robots (Diff-drive, Omni, Ackermann), Quadrupeds (Unitree Go2, DeepRobotics Lite3), Humanoids |
🚀 Quick Start
1. Build Docker Environment
git clone https://github.com/dfl-rlab/dddmr_navigation.git
cd dddmr_navigation/dddmr_docker/docker_file
./build.bash2. Try the Go2 Simulation in Gazebo
Follow our Beginner Guide to launch the Go2 quadruped in Gazebo:
# Terminal 1: Gazebo Simulation
cd ~/dddmr_navigation/src/dddmr_beginner_guide && ./run_x64_gazebo.bash
# Terminal 2: 3D Navigation
cd ~/dddmr_navigation/src/dddmr_beginner_guide && ./run_x64_navigation.bash
ros2 launch p2p_move_base go2_localization.launch3. Deploy on a Real Robot
Consult the Real Robot Deployment Guide for step-by-step TF configuration, online mapping, and 3D goal execution.
📚 Citation
If you use dddmr_navigation in your research or commercial projects, please cite:
@software{dddmr_navigation_dfl_rlab_2025,
author = {CM, PS, Tarek Taha},
title = {dddmr_navigation: 3D Mobile Robot Navigation},
url = {https://github.com/dfl-rlab/dddmr_navigation},
year = {2025}
}🙏 Acknowledgements & Community
The DDDMR Navigation Stack stands on the shoulders of giants. We express our deepest gratitude to the open-source robotics community.
Most importantly, thank you to all our early testers, collaborators, and contributors over the past five years who reported issues, shared rosbag datasets, and helped refine every algorithm in this stack!