Loading experience
0%

Azarel Robotics Engineering

Scroll
Trained by every robot Built for any robot Trained by every robot Built for any robot
Introducing

Sea trials, before the sea.

Product

Leviathan

Leviathan is a development tool for marine robotics — simulate the vehicle, the sensors and the sea, train the autonomy, and grade it before it ever gets wet. From survey and inspection to hull cleaning.

Documentation

Leviathan desktop app, Coral Reef scene — vehicles and environments, ready to run. Leviathan desktop app, Shipwreck scene — six-degree-of-freedom marine physics. Leviathan desktop app, Sensors panel with a live USBL acoustic fix — a complete marine sensor suite. Leviathan desktop app, Chart dashboard panel — train the autonomy, then grade it. Leviathan desktop app, LiDAR dashboard panel with class labels — labeled data, generated rather than collected. Leviathan desktop app, Dam / Reservoir scene — out of the simulator and onto the vehicle.
Leviathan  Viewport scene  Coral Reef

Every run starts at the surface.

Nine environments · one of them below

Environment

Under ice

One of nine environments that ship with Leviathan, and the only one with a ceiling as well as a floor. Every frame below is the cockpit viewport, unretouched.

Leviathan cockpit view of the ice environment: surface and keel in one frame. Leviathan cockpit view of the ice environment: a fin-actuated auv on transit. Leviathan cockpit view of the ice environment: same water, different vehicle. Leviathan cockpit view of the ice environment: looking up at the canopy. Leviathan cockpit view of the ice environment: when the camera stops helping.
preset  ice 01 / 05

Surface and keel in one frame

The air–water interface is part of the model, so the floes above the line and the keels below it belong to the same scene rather than two stitched ones.

Physics engine

FAB Engine

A physics and rendering core purpose-built for the ocean: six-degree-of-freedom hydrodynamics, physically grounded underwater optics, and physically grounded acoustics, engineered together as one system.

WebGL isn't available in this browser, so the demo can't render here.

How FAB Engine Works

Authoring

CAD mesh
Vehicle model
Scenarios
Bathymetry
Materials

FAB Core

Coefficients

Panel solver
Drag
Hull repair
Provenance

Dynamics

6-DOF core
Seakeeping
Currents
Manipulators

Perception

Optical
Acoustic
Sensors
Neural fields

Consumers

RL training
Evaluation
Datasets
ROS 2, HIL
Cockpit

Verification

Analytic oracle
Closed forms
Vehicle data
Field captures
Graded ledger

Geometry-derived hydrodynamics

Hydrodynamic coefficients are computed directly from vehicle geometry. Added mass is resolved on the wetted surface via a boundary-element solver, accurate to 0.1% against analytic reference values.

Validated against published vehicle data

Evaluated on a documented AUV hull, added mass resolves within 0.25% and drag within 2.1% of tow-tank-verified coefficients.

Physically grounded underwater rendering

Novel-view synthesis trained and evaluated on captured underwater imagery, achieving 31.4 dB on held-out views, with explicit recovery of medium attenuation and backscatter.

High-fidelity acoustic simulation

Neural acoustic reconstruction, multipath propagation, range-dependent reverberation and statistically correct speckle. Seabed backscatter is derived from material impedance.

Cross-tier verification

The analytic core serves as the reference implementation. The GPU tier is verified against it term-by-term to bit-level agreement.

Production-scale throughput

1.25 million environment-steps per second on a single consumer GPU, inclusive of observation, reward and reset computation.

Complete six-degree-of-freedom dynamics

Radiation memory, wave excitation, viscous roll damping, and articulated manipulators with joint limits and contact resolution.

Documented operating boundaries

A published limitations reference specifying model scope, exclusions, and the validation basis for each result.

Product

XBOT

XBOT is a robotics-specialized coding model by Azarel Robotics. Ask it for controllers, kinematics, MuJoCo setups or training loops — anything on the stack.

Open XBOT

XBOT runs as a chat in the browser and connects to your own XBOT server, an OpenAI-compatible vLLM endpoint. Settings stay in memory for the session; nothing is stored.

Ask it for

  • An OSC controllerOperational-space control for a 7-DOF Franka Panda in MuJoCo.
  • Nullspace projectionAn explanation for redundant arms, with a short code example.
  • MuJoCo MenagerieLoad the Panda and print its joint limits.
  • A training-loop reviewReview a QLoRA training loop for a code model and suggest fixes.
Blog

Notes from
the build.

All posts →
Field

What underwater robots can do after a flood like Nepal’s, and what they cannot

A glacier collapse on 26 August sent a debris flow nearly 100 km down the Trishuli. No robot was going to help that morning. Flooded powerhouse tunnels, new barrier lakes and scoured bridge piers are a different question.

 Leviathan

You cannot rehearse a flood, so you have to simulate one

No visibility, no map, live current and debris where the seabed should be — and no way to practise in it beforehand. What a simulator owes a team before anyone sends a vehicle in.

 Defence

Looking for one thing in a hundred square kilometres

Mine countermeasures and munitions survey are search problems with a brutal prior. The bottleneck is not the sonar — it is that nobody can give you enough labelled examples of what you are looking for.

 Science

Under ice, surfacing is not an option

One vehicle, one season, no recovery if it goes wrong. Nowhere in marine robotics is practice more valuable, or more impossible to get.
FAQ

Ask less. Know more.

The things marine robotics teams ask first.

What is Leviathan?
+
A development environment for marine robotics. It simulates the vehicle, its sensors and the water around them, lets you train autonomy against that, and grades the result before anything is lowered over the side. Survey, inspection, docking, hull cleaning and under-ice transit are all in scope.
What do I need to run it?
+
The analytic tier runs on an ordinary laptop with no GPU, which is where most day-to-day work happens. The GPU tier is there when you need throughput, and dispatches to a cluster when you have one.
Can I use my own vehicle and my own site?
+
Yes. Eight vehicle presets and nine environments ship with it, but you can bring your own hull mesh and the hydrodynamic coefficients are derived from the geometry rather than hand-tuned. Scenes and vehicles move in and out through OpenUSD, URDF, SDFormat and MJCF.
Does it talk to the stack we already run?
+
There is an ArduSub SITL physics backend, a ROS 2 bridge, and rosbag and pose-log replay straight into the viewport, so a recorded dive can be watched back against the simulated one.
Which sensors are modelled?
+
Twenty-five classes, including a coherent multibeam model that produces side lobes, range ambiguity and speckle the way a real echosounder does, alongside DVL, USBL, LBL, CTD, ADCP, optical and event cameras, and acoustic comms with a full link budget.
How do I know a trained policy will hold up in the water?
+
You do not know from a single run, which is why the benchmark is frozen and versioned. Sixteen tasks are scored per seed with control-stage attribution, so a result can be reproduced and argued with rather than simply asserted. The operating boundaries of the physics are documented rather than implied.
Can I try it?
+
Leviathan is in active development and not generally available yet. If your team has a vehicle and a problem you would rather not debug at sea, write to hello@azarel.com and tell us about it.

One brain for every robot.

Get in touch →