A multi-container Docker-based UAV navigation system integrating LiDAR–Inertial SLAM, real-time trajectory planning, PX4 flight control, and custom ROS1/ROS2 communication, developed as part of a Master’s thesis at Chalmers University of Technology:
Navigation and Localization for Railway Inspection Drone in GPS-denied Environments
This repository contains:
- A complete modular autonomous UAV pipeline
- System-level evaluation and failure analysis
- Early-stage exploratory prototypes on vision-based self-supervised learning
The project bridges AI perception, robotics system integration, and real-time control.
Design and evaluate a full-stack autonomous UAV navigation system capable of operating in GPS-denied, cluttered environments.
The system focuses on:
- Robust LiDAR-based localization
- Dynamic trajectory replanning
- Real-time flight control integration
- Cross-version ROS interoperability
- System-level error propagation analysis
Beyond implementation, the project investigates structural limitations of modular pipelines and explores feasibility of end-to-end learning alternatives.
High-level closed-loop pipeline:
Gazebo Harmonic (Simulation)
↓
ROS2 (IMU + LiDAR topics)
↓
Custom UDP Bridge (ROS2 → ROS1)
↓
FASTER-LIO (LiDAR–Inertial SLAM)
↓
Fast-Planner (Topological Replanning)
↓
Custom UDP Bridge (ROS1 → ROS2)
↓
PX4 EKF2 (State Fusion)
↓
Offboard Control Execution
Deployed using 7 Docker containers, fully isolated and networked via Docker Compose.
While integrating established open-source systems (PX4, Faster-LIO, Fast-Planner), the following components were designed and implemented in this project:
- ROS1 Noetic + ROS2 Jazzy coexistence
- Deterministic Docker networking
- Version isolation
- Reproducible build environment
Developed a custom UDP-based bridge instead of ros1_bridge due to performance and determinism requirements.
Features:
- 2–4 ms average latency (localhost IPC)
- Timestamp preservation
- Explicit serialization control
- Deterministic data flow
- Modular design for future ROS2 migration
- SLAM → Planner → PX4 closed-loop execution
- EKF2 configuration and sensor fusion tuning
- ENU ↔ NED quaternion transformation handling
- Covariance scaling for vertical uncertainty
- Real-time replanning under 20 Hz LiDAR constraint
Performed isolation and analysis of failure sources:
- SLAM localization drift
- Flight controller tracking error
- Planner geometric infeasibility
- LiDAR vertical blind zone
- CPU load profiling
This project emphasizes cross-module debugging rather than isolated algorithm performance.
Dense forest simulation (80 m × 20 m), GPS denied.
| Metric | Value |
|---|---|
| Mission success rate | 77% |
| Pose RMSE (mean) | 2.11 m |
| Rotation RMSE (mean) | 8.46° |
| Trajectory tracking RMSE | 0.21 m |
Identified primary failure sources:
- SLAM drift under dense vegetation
- PX4 tracking deviations during aggressive replanning
- Planner discretization limitations
- LiDAR vertical blind zone during descent
These analyses were conducted to understand structural error propagation in modular autonomy pipelines .
- ROS1 Noetic
- ROS2 Jazzy
- PX4 1.15
- Gazebo Harmonic
- Faster-LIO
- Fast-Planner
- Docker / Docker Compose
- UDP Socket Communication
- C++ / Python
In addition to the modular SLAM pipeline, this repository includes exploratory notebooks investigating vision-based end-to-end learning feasibility:
next-token-prediction.ipynbself-distillation.ipynb
These prototypes explore:
- DINO-based patch embeddings
- Temporal feature forecasting
- Tile-based self-distillation objectives
- Representation collapse behavior
- Structural limitations of naive next-token prediction for video
The experiments revealed:
- Trivial copy solutions in feature forecasting
- Instability in tile-based SSL without multi-view constraints
- Simulation realism limitations for vision-based RL
These findings informed the thesis conclusion that pure end-to-end vision approaches remain structurally constrained in current UAV settings .
The goal of these notebooks is feasibility analysis rather than production-ready learning systems.
- Docker
- NVIDIA Container Toolkit (optional)
- X11 forwarding
git clone https://github.com/method-implanta/RailwayInspection.git
cd RailwayInspection
docker compose build
xhost +local:root
docker network create ros-net
docker compose updocker exec -it gazebo-px4-container /bin/bash
ros2 run offboard_control offboard_control_node- Cross-version ROS interoperability
- Deterministic IPC under multi-container setup
- Real-time SLAM + replanning constraints
- Frame convention mismatch (ENU ↔ NED)
- Sensor update rate bottlenecks (20 Hz LiDAR)
- Modular pipeline abstraction loss
This project demonstrates:
- Full-stack robotics system integration
- Real-time perception–planning–control pipeline design
- Cross-process communication engineering
- SLAM-based autonomy
- Learning-based feasibility analysis
- Structured debugging across heterogeneous modules
It bridges:
AI perception
- Robotics middleware
- Motion planning
- Flight control
- Systems engineering
This system supports the Master’s thesis:
Navigation and Localization for Railway Inspection Drone in GPS-denied Environments
Chalmers University of Technology, 2026
The thesis additionally evaluates limitations of world models, self-supervised learning, and vision-based RL compared to modular SLAM pipelines.
Guanfei Wang
Master’s Thesis – Complex Adaptive Systems
Chalmers University of Technology