NAVI-X — Navigation and Autonomous Visual Intelligence for eXploration
NAVI-X: Enabling intelligent, GPS-independent drone navigation for precise and reliable flight where GPS fails.
- Problem, approach and result explained
- Public source code
- Answered the full brief· 13 of 13
- Expert reviewed· 6 reviews
Overview
The problem
- WHO FACES THIS PROBLEM?
- UAV operators working in indoor facilities and covered industrial environments.
- Inspection teams operating in tunnels and GPS-obstructed areas.
- Organizations requiring reliable autonomous drone navigation where GPS signals are unavailable or unreliable.
- WHY DOES IT MATTER?
- GPS signals become unreliable or unavailable in obstructed environments.
- Conventional GPS-dependent drones struggle with accurate positioning.
- Position drift can cause navigation errors and unsafe flight paths.
- Reliable autonomous navigation and precision landing remain important challenges.
- WHAT DID WE SET OUT TO DO?
- Develop a GPS-independent 3D UAV navigation system.
- Estimate position using onboard motion sensors, camera vision and ground tracking.
- Plan and follow a predefined safe navigation corridor.
- Correct accumulated positioning errors through sensor fusion.
- Enable reliable autonomous navigation and precision landing.
The approach
What did we build and how? We developed a physical UAV prototype, a virtual ground station for live telemetry, eight neural network models, and an ArUco/AprilTag-based precision-landing system. Our approach combines AI-based navigation, onboard sensing, computer vision and ground monitoring.
What was the hardest decision? Choosing a practical navigation architecture that balances accuracy, computational requirements and flight safety while operating without GPS.
What did we choose not to do, and why? We avoided depending entirely on GPS or a single navigation sensor. We also avoided claiming fully autonomous flight before completing hardware validation, prioritizing modular development and controlled testing.
The result
Results Achieved:
- Trained 8 neural network models; the best MLP achieved 93% accuracy with 28,740 trainable parameters.
- Assembled a complete physical UAV.
- Developed a virtual ground station with live telemetry monitoring.
- Integrated ArUco/AprilTag-based precision-landing capabilities.
- Partially developed a laser-based charging station.
Measurement and Comparison: Model performance was evaluated using neural-network accuracy. The 93% result represents the best model's reported accuracy, not flight-navigation accuracy. No verified baseline comparison or quantitative real-world navigation results are available yet.
Current Outcome: Hardware flight stability and complete autonomous GPS-denied navigation remain under validation.
Reflection
What would we do differently?
- Begin hardware integration and testing earlier.
- Validate individual modules before full-system integration.
- Conduct more systematic flight and sensor calibration tests.
What is unfinished or broken?
- Drone stability issue causing unwanted tilting.
- Full GPS-denied autonomous navigation and flight validation.
- Complete precision-landing validation.
- Fully operational laser charging station.
What did we learn?
- Hardware integration is more challenging than simulation alone.
- Sensor calibration and system reliability are critical for UAV navigation.
- AI model accuracy does not necessarily guarantee real-world performance.
- Modular development and continuous testing are essential for reliable autonomous systems.
Submitted to 3D Autonomous Path Planning
- Your idea in one line
NAVI-X 1. is GPS-denied UAV Using ROS 2 and PX4 that flies a 3D corridor using a deep reinforcement learning neural network as its primary pilot, with an EKF fusing IMU, camera, and ground-tracking data, and a PID override as the emergency airbag if the NN fails.
- What will be working by the finale?
By the finale, we will deliver a working simulation of a GPS-denied autonomous UAV that: 1. Flies a 3D corridor using a trained deep reinforcement learning (PPO) neural network as the primary autonomy layer, taking fused state and camera images as input. 2. Localizes accurately without GPS using a confidence-aware EKF that fuses IMU dead reckoning, visual odometry from the onboard camera, and ground-tracker position fixes. 3. Enforces corridor safety through a geometric monitor (D_c ≤ R_c) that validates every NN action before it reaches the motors. 4. Fails safely via a PID override that activates if the NN produces invalid output, the EKF diverges, or the corridor is breached — keeping the drone stable until recovery or return-to-launch. 5. Performs autonomous precision landing within the designated pad radius (E_L ≤ R_landing). 6. Is validated quantitatively through an ablation study (IMU-only vs. IMU+Ground vs. IMU+Ground+Camera) and Monte Carlo runs across sensor noise, dropouts, and wind disturbances — reporting position RMSE, corridor deviation, landing error, and success rate.
- Tech stack and hardware you plan to use
Flight Control and Simulation: - PX4 Autopilot — autonomous flight control and vehicle-state estimation. - Gazebo Harmonic — 3D simulation of the UAV, obstacles, sensors, and operating environment. - X500 Quadcopter Model — simulated UAV platform. - Simulation & Middleware (already set up): · PX4 Autopilot — flight stack, running on Pop!_OS 24.04 · ROS 2 Jazzy — inter-module middleware and topic communication · Gazebo Harmonic — 3D physics simulation · Micro XRCE-DDS Agent — PX4 ↔ ROS 2 bridge · X500 quadcopter — simulated airframe in SITL Autonomy Stack (in development): · EKF / INS Fusion — 15-state estimator fusing IMU, visual odometry, and ground tracking, with innovation gating and confidence-aware weighting · Visual Odometry — ORB feature detection + KLT optical flow via OpenCV · Neural Network — Deep Reinforcement Learning (PPO) policy implemented in PyTorch / Stable-Baselines3, with a CNN for camera input and MLP for state input (~200–500k parameters) · Safety Layer — geometric corridor monitor (D_c vs R_c) + safety arbiter + PID override controller · Training Environment — Gymnasium wrapper around PX4 SITL / Gazebo, with domain randomization; optionally migrating to Isaac Lab or AirSim for photorealistic training Deployment Target: · Companion computer — pico pi, raspberry pi · Inference — PyTorch → ONNX → TensorRT for real-time policy execution · Flight controller — Pixhawk-class hardware The neural network is our primary autonomy layer — it takes fused state and camera images as input and outputs flight commands. The EKF provides localization; the geometric corridor and PID form the safety envelope around the NN.
- What have you built or tested so far?
* Set up the PX4 Autopilot development environment on Pop!_OS 24.04. * Built the PX4 Software-in-the-Loop (SITL) simulation for the X500 quadcopter. * Installed and configured ROS 2 Jazzy and Gazebo Harmonic. * Built the Micro XRCE-DDS Agent to support communication between PX4 and ROS 2. * Started the simulated drone and verified that the PX4 process and DDS Agent can run. Current status: The PX4–ROS 2 topic connection is still being debugged. Autonomous GPS-denied navigation, obstacle avoidance, and quantitative performance evaluation remain to be implemented and tested. Next steps: 1. Establish and verify the PX4–ROS 2 connection. 2. Integrate simulated sensor data. 3. Implement local mapping or obstacle representation. 4. Develop and test a path-planning algorithm. 5. Evaluate navigation success, path length, obstacle clearance, and computation time.
- Pitch deck (PDF)
- Pitch deck (PDF)PDF
- GitHub repository
- github.com/psmahanambrata/iceberg_gps_denied_UAV
- What's working now?
Working: - Complete physical drone assembled. - Virtual mini ground station developed with live drone parameter monitoring. - Eight neural network models trained; the best MLP achieved 93% accuracy with 28,740 trainable parameters. - ArUco markers and AprilTags integrated for precision-landing development. - Earlier ROS 2, PX4 and simulation-based navigation development completed at the software-foundation level. Partly Working: - Physical UAV integration and testing are ongoing. A minor hardware issue is causing drone tilting and is under troubleshooting. - Marker-based precision landing is integrated but requires further flight validation. - Ground-station monitoring is available, while complete autonomous navigation integration remains under development. Not Yet Achieved: - Fully validated GPS-denied autonomous 3D corridor flight. - Reliable real-world position estimation and continuous drift correction. - Complete end-to-end autonomous navigation and landing demonstration. - Operational laser-based drone charging; only a preliminary charging-station prototype has been developed. We have progressed from simulation and software development to a physical UAV prototype, AI model development and supporting ground infrastructure. Full system-level validation remains our next major milestone.
- GitHub repository
- github.com/ghqadir7827-droid/bput_uav_iceberg
- Deck or report (PDF)
- Deck or report (PDF)PDF
AI tools used
we took it as reference for our data sets, to train our models with higher accuracy, to understand problems we couldn't configure.