Scribd
Upload
39 views23 pages

Stonefish Simulation

Uploaded by

Paco Martinez
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
39 views23 pages

Stonefish Simulation

Uploaded by

Paco Martinez
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

www.sfs.com.au www.missionsystems.com.

au

Distributed MCM Simulation


with MOOS-IvP
David Battle
([email protected])

10th Aug 2022


What is S4?
S4 is a simulation system tailored to the requirements
of autonomous mine countermeasures.

Introduction
MINTACS user interface
Constructs a synthetic environment within UE4

The MINTACS mine warfare tactical


planning suite has been interfaced to
the UE4 game engine to create a
high-performance simulation and
visualisation system.
We call this S4.
www.sfs.com.au
S4 Simulator architecture
Full prototype
S4 Simulator architecture
Fast-data generation subset

UE4 only required to initialise


the environment and can run
headless or be shutdown
thereafter.
REMUS 100 running near the shoreline
Bluefin 9 in minefield
AQS-24 towbody during a turn
Tow platform in sea state 2
Use cases
1. Automatic Target Recognition (ATR) Training
2. Operational Mission Planning
3. Operational Mission Analytics
4. Mine Warfare Personnel Training
S4 Prototype: These have now been broken down into four distinct modes of simulation:
• Mission-Mode (previously time-warp), in which platform dynamics are
Modes and use cases applied in an accelerated manner along a specific mission-track in a
selected region of interest.
• REA-Mode experimental mode in which region of interest is evaluated
for the detectability of specific mine threats using a specific sonar.
• RMA-Mode Another experimental mode in which a particular vehicle
track is evaluated for the detectability of specific mine threats in a
specific environment.
• FastDataGen-Mode (previously ludicrous-mode), in which a
disembodied sonar teleports around without vehicle dynamics to
generate synthetic data for ATR training
Headline use case:
Data generation for synthetic ATR training
Target simulation
S4 can produce infinite amounts of synthetic sonar data!

Manta mine Rockan mine IHO cube Stonefish mine Buoyant mine
Target simulation cont’d
Mine burial and accretion simulation
Target simulation cont’d
Real vs synthesized data
Synthetic sonar data
(Marine Sonics 900 kHz – one side)
Synthetic sonar data
(Sonardyne SOLSTICE 750 kHz – one side)
Fast data generation
The ATR training workhorse
Headline
Achievements: ● Place any mine type on any sediment
Add any clutter from the real world
ATR Training Pipeline ●

● Create any approach vector desired


● Generate 10,000+ mine images over
night (on a single 3090 GPU)
● Operates independently of MINTACS
ATR Training Pipeline
Adding synthetic clutter to reduce false alarms
• The central thesis of this project was that optimal ATR
detection performance requires training in virtual
environments with statistics as similar to reality as
possible.
• As a corollary, the addition of clutter objects to the
training data which are rejected as targets should then
reduce the false alarm rate.
• In trials to date, we have seen false alarms drop from
over 300 in real mission data to under 100 simply by
adding schools of fish into the simulated training data!
ATR Training Pipeline cont’d
Training example with data augmentation (YoloV5s)
Synthetic data: 8221 Images:
4654 stonefish, 3034 manta,
2609 rockan, 1786 buoyant
mines
Synthetic + real data: 8221
Images + 654 with 154 MLOs
ATR Training Pipeline cont’d
ATR Challenge ROC Curves (Marine Sonics 900 kHz)

YoloV5 trained by S4 with ATR Challenge + synthetic data Haar Cascade Detector trained with ATR Challenge data
86.9% TPR with 4751 False Alarms 70% with 3500 False alarms and 14,000 False alarms for 80% TPR
MOOS-IvP feature wish list
Things that might make life easier*

• Runtime warp control


• Better shutdown control
Conclusions: • Configurable message queue depth
• Services
• Name spaces
• Encryption and access control
• Config syntax highlighting
• MOOSDB reset

* Some of these suggestions were the result of a survey of our programmers,


some of whom have not been using MOOS for very long. Some, however
come from experienced MOOS practitioners.

OFFICIAL
• We have demonstrated a functioning prototype of the S4
MCM simulation system
Conclusions: • S4 is based on MOOS-IvP because of its small codebase
and limited dependencies, which simplified its integration
into MINTACS
• In the course of demonstrating S4, we have pushed
MOOS-IvP to the limit and it didn’t break!
• We have demonstrated that synthetic sonar data can be
used to train ML-based ATR systems
• The next version of S4 will be demonstrated at Euronaval
later this year.
Questions?

You might also like