AQUA Stack Overview¶

A comprehensive guide to understanding the AQUA Stack autonomy platform for underwater vehicles.

What is AQUA Stack?¶

AQUA Stack is a complete software platform for autonomous underwater vehicles (AUVs) that provides:

GPS-denied navigation using visual SLAM
Autonomous mission execution with behavior trees
Real-time path planning and obstacle avoidance
Vehicle-agnostic interface through ArduSub integration
Simulation tools for safe mission testing

The Challenge of Underwater Autonomy¶

Autonomous underwater operations present unique challenges:

No GPS¶

Unlike aerial or ground robots, underwater vehicles cannot use GPS for localization. Water blocks GPS signals beyond a few meters depth.

Limited Communication¶

Underwater communication is restricted to: - Acoustic: Low bandwidth, high latency - Tether: Limited range, drag effects - None: Fully autonomous operation required

Dynamic Environment¶

Currents: Constant disturbances to planned paths
Poor visibility: Turbidity reduces camera range
Variable lighting: Depth and time-of-day variations

Safety Critical¶

Equipment is expensive (>$50K typical)
Recovery is difficult if lost
Environmental risks (entanglement, collision)

How AQUA Stack Solves These Challenges¶

1. Visual SLAM for Localization¶

Instead of GPS, AQUA Stack uses Visual SLAM (Simultaneous Localization and Mapping):

graph LR
    A[Camera Images] --> B[Feature Detection]
    B --> C[Feature Tracking]
    C --> D[Pose Estimation]
    D --> E[Map Update]
    E --> F[Loop Closure]
    F --> D

    style D fill:#4CAF50
    style F fill:#FF9800

How it works: 1. Extract visual features from camera images 2. Track features across frames 3. Estimate camera (vehicle) pose from feature motion 4. Build 3D map of environment 5. Detect when vehicle returns to known areas (loop closure) 6. Correct accumulated drift

Result: Accurate position estimation without GPS

2. Behavior Trees for Mission Control¶

AQUA Stack uses behavior trees for flexible, reactive mission execution:

Mission: Survey Pipeline
├─ Sequence
│  ├─ GoToWaypoint(start)
│  ├─ Repeat(10)
│  │  └─ Sequence
│  │     ├─ FollowPath(pipeline)
│  │     ├─ TakeSurveyPhoto
│  │     └─ AdvanceAlongPath
│  └─ ReturnToHome
└─ Fallback (if above fails)
   └─ EmergencyS urface

Advantages: - Composable: Build complex behaviors from simple primitives - Reactive: Respond to changing conditions - Debuggable: Clear structure for troubleshooting - Reusable: Share behaviors across missions

3. Adaptive Path Planning¶

AQUA Stack continuously replans based on:

Currents detected: Compensate for water flow
Obstacles discovered: Reroute around hazards
Energy remaining: Adjust mission to ensure safe return
SLAM confidence: Fall back to safer modes if needed

4. Proven Vehicle Integration¶

ArduSub provides the low-level control interface:

┌────────────────────────────────────────┐
│          AQUA Stack                     │
│  (High-level Navigation & Planning)    │
└─────────────────┬──────────────────────┘
                  │ MAVLink
┌─────────────────┴──────────────────────┐
│          ArduSub Autopilot              │
│  (Stabilization, Motor Control)        │
└─────────────────┬──────────────────────┘
                  │ PWM/Serial
┌─────────────────┴──────────────────────┐
│    Vehicle Hardware (Thrusters, ESCs)  │
└────────────────────────────────────────┘

Benefits: - Proven autopilot with thousands of flight hours - Safety features (failsafes, arming checks) - Standard interface (MAVLink protocol) - Large community and support

Architecture Deep Dive¶

Component Layers¶

AQUA Stack is organized into distinct layers:

Layer 1: Perception¶

Input: Raw sensor data (images, IMU, depth) Processing: - Camera calibration and undistortion - Feature detection (ORB, SIFT, etc.) - IMU preintegration - Sensor synchronization

Output: Processed observations for SLAM

Layer 2: Localization¶

Input: Processed observations Processing: - Visual odometry (frame-to-frame tracking) - Local mapping (sliding window optimization) - Loop closure detection - Global pose graph optimization

Output: Vehicle pose estimate + confidence

Layer 3: Planning¶

Input: Current pose, mission waypoints, map Processing: - Global path planning (A, RRT) - Local trajectory optimization - Collision checking - Current compensation

Output: Desired trajectory

Layer 4: Control¶

Input: Desired trajectory, current state Processing: - Path following (pure pursuit, LQR, MPC) - Behavior tree evaluation - Safety checks - Failsafe triggers

Output: Vehicle commands (velocity, heading, depth)

Layer 5: Actuation¶

Input: Vehicle commands Processing: - ArduSub MAVLink interface - Command translation - Telemetry feedback

Output: Motor PWM commands

Data Flow¶

Sensors → Perception → SLAM → World State
                               ↓
Mission Plan → Planning → Trajectory → Control → Actuation
                                                     ↓
                                                  Vehicle

Coordinate Frames¶

AQUA Stack uses several coordinate frames:

World Frame (W)¶

Origin: Mission start location
Orientation: North-East-Down (NED)
Usage: Global planning, waypoint definition

Body Frame (B)¶

Origin: Vehicle center of mass
Orientation: Forward-Right-Down (FRD)
Usage: Vehicle dynamics, control

Camera Frame (C)¶

Origin: Camera optical center
Orientation: Forward-Left-Up (FLU) or as calibrated
Usage: Visual observations, SLAM

Map Frame (M)¶

Origin: First keyframe in SLAM map
Orientation: Arbitrary (typically aligned with world at start)
Usage: SLAM, local navigation

Transform chain: W → M → B → C

Performance Characteristics¶

Localization Accuracy¶

Environment	Drift Rate	Loop Closure	Notes
Clear water, good features	<1% distance traveled	Yes	Optimal conditions
Murky water, few features	2-5% distance traveled	Rare	Challenging
Structured (pipeline, wall)	<0.5% distance traveled	Frequent	Ideal

Computational Requirements¶

Component	CPU Usage	Memory	Notes
Visual SLAM	40-60%	500MB	Single core on Jetson Nano
Planning	10-20%	100MB	Depends on map size
Control	5-10%	50MB	Realtime constraint
Total	~70%	~650MB	With margin for other tasks

Mission Duration¶

Typical mission times with AQUA Stack:

Setup & calibration: 10-15 minutes
Mission execution: Limited by vehicle battery (1-4 hours typical)
Data download & processing: 5-10 minutes

Integration Points¶

AQUA Stack interfaces with:

Inputs¶

Cameras: ROS camera drivers or GStreamer pipelines
IMU: Standard IMU drivers (MPU6050, BMI088, etc.)
Depth sensor: MS5837 or similar
DVL (optional): Teledyne, Nortek, LinkQuest
GPS (surface): For initial heading, mission upload

Outputs¶

ArduSub: MAVLink over serial/UDP
Ground station: Telemetry via acoustic modem or tether
Data logging: ROS bags, custom formats
Mission results: Maps, trajectories, photos

Deployment Scenarios¶

Scenario 1: Tethered Operations¶

Setup: Fiber optic or copper tether Benefits: - Real-time telemetry - Human-in-loop control possible - Continuous power

Limitations: - Range limited (100-300m typical) - Drag affects vehicle performance

Scenario 2: Acoustic Communications¶

Setup: Acoustic modem (e.g., WHOI MicroModem) Benefits: - Longer range (1-5km) - Untethered freedom

Limitations: - Very low bandwidth (100-1000 bps) - High latency (seconds) - Expensive ($10K+ for modem)

Scenario 3: Fully Autonomous¶

Setup: Pre-loaded mission, no communication Benefits: - Maximum range (limited only by battery) - No communication infrastructure needed

Limitations: - No real-time monitoring - Must complete mission or return safely - Higher risk

Safety Features¶

AQUA Stack includes multiple safety layers:

Pre-Mission Checks¶

✓ SLAM initialization quality
✓ Sensor health
✓ Battery state
✓ ArduSub arming checks
✓ GPS fix (if surface start)

In-Mission Monitoring¶

🛡️ SLAM confidence tracking
🛡️ Battery consumption monitoring
🛡️ Obstacle detection
🛡️ Depth limits
🛡️ Communication timeout (if applicable)

Failsafe Behaviors¶

Low SLAM confidence → Switch to dead reckoning, head to surface
Low battery → Abort mission, return home
Obstacle detected → Stop, replan, or surface
Communication loss → Continue mission or return based on settings
Depth limit exceeded → Emergency ascent

Comparison with Other Systems¶

Feature	AQUA Stack	Custom Development	Academic SLAM Only
Development time	Days	18-24 months	6-12 months
Mission planning	Included	Build from scratch	Not included
Vehicle integration	Plug-and-play	Custom for each vehicle	Manual integration
Safety features	Production-ready	Must implement	Basic or none
Support	Commercial	None	Academic papers only
Cost	License fee	Engineer time (>$200K)	Free (software only)

Next Steps¶

Now that you understand AQUA Stack:

Check Requirements - Ensure your hardware is compatible
Installation - Set up AQUA Stack
Quick Start - Run your first mission
Tutorials - Detailed walkthroughs