JacopoPan/aerial-autonomy-stack

What it solves

Aerial Autonomy Stack (AAS) provides a unified, "batteries included" software framework for developing, simulating, and deploying autonomous multi-drone systems. It eliminates the friction of setting up complex pipelines by integrating autopilots, simulators, perception algorithms, and deployment tools into a single, Dockerized workflow.

How it works

The project uses a containerized architecture to separate simulation, ground control, and aircraft logic:

• Simulation: Runs PX4 or ArduPilot SITL (Software-In-The-Loop) within Gazebo, supporting faster-than-real-time execution and multiple vehicle types (quadrotors, VTOLs, tailsitters).
• Aircraft Logic: A ROS2-based stack that handles perception (YOLO for vision, KISS-ICP for LiDAR) and control (behavior tree missions and offboard controllers).
• Ground Control: Manages multi-drone coordination via a traffic controller and provides operator supervision through QGroundControl.
• Deployment: Supports Hardware-In-The-Loop (HITL) simulation and direct deployment to NVIDIA Jetson Orin hardware using TensorRT and DeepStream for optimized inference.

Who it’s for

Robotics engineers and researchers focusing on aerial autonomy, multi-agent drone coordination, and perception-based flight who need a seamless transition from simulation to real-world NVIDIA Jetson hardware.

Highlights

• Autopilot Agnostic: Compatible with both PX4 and ArduPilot.
• High-Performance Simulation: Supports faster-than-real-time, multi-instance simulation and a Gymnasium-compatible RL environment.
• Perception Ready: Integrated YOLO (ONNX GPU) and LiDAR odometry (KISS-ICP).
• Deployment Optimized: Dockerized images specifically for NVIDIA Jetson (arm64) with TensorRT and DeepStream support.
• Multi-Vehicle Coordination: Includes a Zenoh-based ROS2 bridge for inter-vehicle communication and a synchronized formation controller.