Training datafor underwaterautonomy

A public-data MVP for underwater autonomy. USIM video, IMU, depth, and control aligned for VLA training.

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

Two live replays

SLAM plus the USIM training-data pipeline.

SLAM replay

Raw · SAM2 · ORB · VO
01Raw Stream
02Segmentation · SAM2
03SLAM Features · ORB
043D Map · Visual Odometry
00:00 / 00:00
$ loading replay artifacts…

Underwater AI lacks training data

Underwater data is mostly mapping data. Policies need perception plus action.

0
Public real-world VLA datasets
Terrestrial robots have them. Underwater robots do not.
25 hrs
Synthetic training data (USIM)
Useful for prototyping; still not real ocean data.
2,275
USIM trajectories
Public simulated BlueROV2 runs for MVP validation.
4
USIM signals per timestep
Video, IMU, depth, and control.

Underwater datasets today

Capture what pilots already do

Validate on USIM now. Design the same row contract for future ROV black-box traces.

Static dashboard first; real telemetry stays separate until a trace exists.

View demo
01
Passive capture

A topside tap records MAVLink traffic without modifying the ROV or interrupting the pilot.

MAVLinkArduSubFathom-X
02
Align and normalize

Streams become uniform rows with scaled PWM commands and consistent action vectors.

SyncNormalizeJSON
03
Auto-label and chunk

Physics labels and ACT-style action chunks export to static JSON first.

HydrodynamicsACTExport

Ingest → align → label → export

Raw ROV streams in. Model-ready chunks out.

I
Action chunking

Smooth control sequences

Exports smoothed future-action chunks as fixed [k, 6] tensors.

II
Cross-modal roadmap

Sydney Harbour after the USIM spine

Add paired optical and Oculus FLS sonar after the static USIM demo works.

III
Physics-derived labels

No human annotator for hydrodynamics

Derives labels like fighting_current from command vs. IMU response.

MVP scope from idea.md

Prove the tokenization path on public USIM data. Keep the dashboard static. Add real MAVLink and sonar only when the source traces exist.

USIM spineStatic dashboardAction chunksSonar later
No backend for v1JSON before ParquetReal logs are the product

Who needs this data

One dataset format for defense, inspection, and research teams.

GPS-denied UUV navigation

Policies for current, turbidity, and acoustic sensing.

Offshore inspection & IRM

Resident autonomy needs better pilot-derived data.

Marine robotics research

Benchmarks for policies, not just perception.

Data and policy, not another OS

Incumbents sell tools. We build the missing training layer.

Greensea IQ · OPENSEA

Navigation OS and sensor integration. A VLA policy can sit above it.

SeeByte · SeeTrack

Mission planning says where to go. Policies handle how to move.

Classical CV stacks

Rules miss edge cases; pilot data captures them.

Work with Aronnax Labs

Built from the Bow Capital x UCSD StartBlue MVP blueprint. Open for pilot conversations and data access.

A
Defense & UUV teams

Teams building attritable underwater platforms that need robust autonomy software.

B
Offshore operators

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

C
ROV fleet partners

Operators who can host passive capture hardware during routine missions.

D
Research labs

Marine and oceanographic groups benchmarking underwater autonomy, not just mapping.

Aronnax Labs · UCSD StartBlue

contact@aronnaxlab.ai
LocationLa Jolla, CA