Embedded Files
What we're working with!
What we're working with!
AquamaRUM's navigation and control systems ensure reliable task execution using a robust Python-based tech stack. The architecture comprises two subsystems: vision and control. The control system communicates with the Ground Control Station (GCS) to publish telemetry and system status, manages the ASV's operational status, interfaces with the flight controller, and initiates movement subprocesses. The ASV operates in two modes: AI/Pathfinding and Missions. Pathfinding mode uses dynamic decision trees, implemented with Python's Py-Trees library, to generate navigation commands based on vision pipeline data. These customizable trees are task-specific and modular. Missions’ mode executes preloaded instructions in a custom scripting language designed for the vessel's operations.
AquamaRUM's navigation and control systems ensure reliable task execution using a robust Python-based tech stack. The architecture comprises two subsystems: vision and control. The control system communicates with the Ground Control Station (GCS) to publish telemetry and system status, manages the ASV's operational status, interfaces with the flight controller, and initiates movement subprocesses. The ASV operates in two modes: AI/Pathfinding and Missions. Pathfinding mode uses dynamic decision trees, implemented with Python's Py-Trees library, to generate navigation commands based on vision pipeline data. These customizable trees are task-specific and modular. Missions’ mode executes preloaded instructions in a custom scripting language designed for the vessel's operations.
How it works!
How it works!
Vision
Vision
The AI and computer vision systems use a vision pipeline for object detection. Data from a RealSense D435 camera is processed by an in-house-trained YOLOv8 model to identify specific object segments. The system minimizes processing overhead by sending frames at a specific rate. Outputs include object type, color, and radial coordinates, enabling real-time navigation.
Simulation
Simulation
AquamaRUM’s simulation framework refines pathfinding algorithms without physical deployment. The environment replicates key subsystems like the vision pipeline and flight controller. Ardupilot SITL simulates vehicle dynamics, while a custom vision simulator creates buoy arrangements based on relative coordinates. The simulator’s lightweight design enables rapid testing and iteration, ensuring seamless transitions from virtual to real-world testing. Additionally, the team is developing a Gazebo environment to expand testing capabilities, allowing for more detailed and realistic simulations of AquamaRUM’s interactions with its surroundings.
Pathfinding
Pathfinding
Once the data of objects seen is retrieved (color, type, location), the team implemented algorithms to determine AquamaRUM’s movement based on vision data. Some of the same algorithms for pathfinding are implemented in similar tasks, like calculating midpoint coordinate between gate buoys, move to a calculated point, move until gate buoys are close enough for midpoint to be calculated, etc.
Simulation
Simulation
AquamaRUM’s simulation framework refines pathfinding algorithms without physical deployment. The environment replicates key subsystems like the vision pipeline and flight controller. Ardupilot SITL simulates vehicle dynamics, while a custom vision simulator creates buoy arrangements based on relative coordinates. The simulator’s lightweight design enables rapid testing and iteration, ensuring seamless transitions from virtual to real-world testing.
AquamaRUM’s simulation framework refines pathfinding algorithms without physical deployment. The environment replicates key subsystems like the vision pipeline and flight controller. Ardupilot SITL simulates vehicle dynamics, while a custom vision simulator creates buoy arrangements based on relative coordinates. The simulator’s lightweight design enables rapid testing and iteration, ensuring seamless transitions from virtual to real-world testing.
What we use
What we use
PyTorch
YoloV8
Ardupilot
MavLink
QGroundControl
Label Studio
Python
NVIDIA
OPENCV
Google Sites
Report abuse