Synthetic data for maritime autonomy

All-weather, all-sea-state perception data for autonomous vessels, USVs, and coastal surveillance. Generate the rare and dangerous scenarios you can never safely collect at sea, labeled to the pixel across EO, IR, and SWIR.

Working model in one week. Zero sea trials.

The data that matters most is the data you cannot safely collect

Maritime AI fails for one reason above all: bad data. Real collection at sea is slow, costly, and weather-dependent, and it skews to calm, common, benign conditions. The cases that decide a mission are exactly the ones you cannot stage: heavy sea states, night, fog, glint, crowded ports, fast inbound small craft.

Stardust generates them on demand. Built on Unreal Engine and a learned rendering stack, it produces photorealistic, fully labeled maritime scenes across the full operational envelope, with EO, IR, and SWIR in registration, and physics-based marine radar in development. Closed-loop simulation then lets your autonomy stack act in the scene and be scored against ground truth, so perception and planning are validated before the first sea trial.

Why real data falls short

• ✕Collecting labeled data at sea is slow, costly, and weather-dependent
• ✕The dangerous cases are exactly the ones you cannot stage: heavy seas, night, fog, glint, near-collisions
• ✕EO alone misses targets that IR and SWIR would catch
• ✕Open-loop data cannot validate autonomy: without closed-loop control, perception and planning failures stay hidden until deployment

Photorealistic data, perfectly labeled

Every sea state and condition Sea state, weather, time of day, sun angle, glint, and visibility across the full operational envelope.

Multi-sensor, registered EO, IR, and SWIR in perfect registration. Physics-based marine radar in development with Ansys.

Pixel-perfect labels and 3D metadata Detection, classification, and segmentation with ground truth on every frame, plus platform pose, object range, and class counts.

Closed-loop simulation Step through scenarios with a sim-in-the-loop API, replay real missions, and run CI batteries against your autonomy stack.

Inside the maritime data

• Neural-rendered sea: Photorealistic harbor and open-water scenes
• Photorealistic, with ground truth: Live platform, environment, and objects-in-range metadata
• Every modality, registered: EO, segmentation, and depth from one scene
• Closed-loop telemetry: Multi-sensor mission replay in Rerun

Explore a real maritime scene, frame by frame

Every Stardust scene exports to Rerun. Pan and zoom the 3D world, scrub the timeline, and toggle sensor streams in registration. Live, interactive, real Stardust output.

What teams build with it

• Collision avoidance: Detect and track vessels and obstacles in all conditions.
• COLREGS scenario testing: Validate behavior against ready-made scenarios that mimic the rules of the road.
• Automatic target recognition: Classify vessels by type across EO and IR.
• Vessel detection and classification: Boats, buoys, launches, and tugs, by range and class.
• Maritime domain awareness: Persistent wide-area surveillance and tracking.
• Port and harbor monitoring: Watch busy, cluttered waterways around the clock.
• USV and ASV patrol: Train and test autonomous surface vessel perception at scale.
• Dark-vessel and small-craft detection: Surface low-signature and fast inbound contacts.

Speaks your domain

The vocabulary, sensors, and benchmarks maritime teams actually use.

“We hit great real-world performance almost immediately. But even more impressive, Bifrost’s 3D metadata let us develop AI capabilities that just are not possible with real data.”

Perception lead, defense technology company valued over $30B

Questions teams ask

How do you train a USV to avoid collisions without sea trials?

Generate thousands of labeled collision and COLREGS scenarios across sea states, weather, and traffic, then close the loop so the autonomy stack acts and is scored against ground truth.

What sensors do you need for maritime ATR?

EO/RGB, thermal IR, and marine radar. Stardust generates EO and IR today with pixel-perfect labels and 3D metadata, with physics-based marine radar in development.

Can you simulate marine radar?

Marine radar is in active development, in partnership with Ansys. It will support configurable carrier frequency, beamwidth, polarization, and antenna pattern for radar system validation.

How do you cover the long tail at sea?

Find the gaps and biases in your real data, then generate targeted synthetic data patches to fill them, around 100x faster than collecting more at sea.

What is closed-loop maritime simulation?

A step-through simulation where your autonomy stack observes synthetic sensor frames and issues controls each timestep, so perception and planning can be regression-tested in CI.

Validate maritime autonomy before the first sea trial

Tell us what you are building and the scenarios you need. We will get you access.