Training datafor underwaterautonomy

Real-world data for underwater autonomy. We capture aligned video, sonar, IMU, and pilot control from ROV operations and turn it into training-ready VLA chunks.

0
public real-world VLA datasets
25h
synthetic data (USIM, industry-wide)
4
modalities per timestep
scroll

Underwater AI lacks training data

Models trained on land or in simulation struggle underwater. Current, turbidity, and radio blackout change the sensing problem entirely. Without aligned vision-action-telemetry datasets, marine autonomy stays rule-based.

0
Public real-world VLA datasets
Terrestrial robotics has nuScenes, DROID, and more. Underwater has none.
25 hrs
Synthetic training data (USIM)
Useful for prototyping, but sim-to-real gaps remain large.
$150K+
Typical ROV inspection day rate
DP vessel + crew makes field data expensive to collect at scale.
4
Modalities we align per timestep
Video, sonar, IMU, and pilot control — timestamped together.

Underwater datasets today

DatasetDomainModalitiesSizeActionsNotes
nuScenes / DROIDTerrestrialVision, LiDAR, controlHundreds of hoursNo fluid dynamics
USIMUnderwaterVision, control25 hrs · 905K framesSynthetic only
SOVISUnderwaterVision, sonar76K framesPerception only
Aronnax (in progress)UnderwaterVision · sonar · IMU · controlPilot deploymentsReal ROV telemetry — pipeline live on USIM today

Capture what pilots already do

Instead of building new vehicles, we record telemetry from commercial ROV missions — inspections, surveys, and maintenance — and process it through a single annotation pipeline.

The pipeline is validated on public USIM simulation data today and designed to run unchanged when real black-box ROV traces arrive.

Open dataset explorer
01

Passive capture

A hardware tap on the ROV topside Ethernet bridge records MAVLink traffic — video, sonar, joystick input, IMU, depth, and pressure — without modifying the vehicle or interrupting the pilot.

MAVLinkArduSubFathom-X
02

Align and normalize

Streams are timestamped and normalized into uniform rows. PWM commands are scaled to a fixed contract so training code sees consistent action vectors across vehicles.

SyncNormalizeParquet
03

Auto-label and chunk

Physics-derived labels (e.g. fighting_current when thrust does not match IMU response) and ACT-style action chunks produce export-ready JSON and Parquet for model training.

HydrodynamicsACTExport

Ingest → align → label → export

Four stages turn raw ROV streams into training-ready chunks. Live demo runs on USIM today.

I

Action chunking

Smooth control sequences

Rather than one thruster command per frame, the pipeline exports ACT-style chunks of future actions with exponential smoothing — fixed [k, 6] tensors ready for imitation learning.

ACTChunkingImitation learning
II

Cross-modal sonar

Vision when you have it, sonar when you do not

Camera frames pair with forward-looking sonar masks in a shared timeline. When optical visibility drops, the same row structure carries sonar frames and detection boxes through export.

FLS sonarMultimodalAlignment
III

Physics-derived labels

No human annotator for hydrodynamics

When a pilot commands forward thrust but IMU acceleration stays near zero, the pipeline tags fighting_current — a subjective pilot reaction turned into an objective training token from RC_IN vs SCALED_IMU.

IMUHydrodynamicsAuto-label

Build command

python -m underwater_vla build --limit 200 --with-sonar
Ingest raw telemetryAlign to uniform rowsDerive VLA labelsChunk actions → JSON / Parquet

Who needs this data

The same aligned telemetry supports defense autonomy, commercial inspection, and academic benchmarking — any team training policies that must work underwater.

GPS-denied UUV navigation

Low-cost underwater vehicles need policies that handle current, turbidity, and acoustic sensing — not just pre-programmed waypoints.

Offshore inspection & IRM

ROV day rates and pilot shortages push operators toward resident autonomy. Better training data is the bottleneck.

Marine robotics research

Labs need aligned multimodal traces — not just perception frames — to benchmark underwater policies.

Data and policy, not another OS

Incumbents sell integration software and mission tools. We focus on the missing layer: aligned training data and learned control policies for underwater vehicles.

Greensea IQ · OPENSEA

Rules-based navigation OS and sensor integration layer. A VLA policy could sit on top of their edge stack as a behavioral layer.

SeeByte · SeeTrack

Mission planning and fleet C2. They handle where to go; a trained policy handles how to move moment-to-moment.

Classical CV stacks

ATR and rule-based obstacle avoidance break on novel debris and zero visibility. Learned policies from pilot data address edge cases rules miss.

Work with Aronnax Lab

We are incubating at UC San Diego StartBlue with Scripps Institution of Oceanography. Reach out for pilot partnerships, data access, or research collaboration.

A

Defense & UUV teams

Teams building attritable underwater platforms that need robust autonomy software.

B

Offshore operators

Inspection and IRM contractors looking to reduce vessel days and pilot load.

C

ROV fleet partners

Operators who can host passive capture hardware during routine missions.

D

Research labs

Oceanographic and marine robotics groups benchmarking underwater policies.

Aronnax Lab · StartBlue × UCSD Scripps

contact@aronnaxlab.ai
LocationLa Jolla, CA
ProgramStartBlue · UC San Diego
StageSeed · data collection