Synthetic data for off-road autonomy

Unstructured-terrain perception across occlusion, dust, and degraded visibility. Generate the terrain and conditions that break perception, before they break it in the field.

Endless terrain, before the field tests you.

Every scene off-road is a new scene

Unstructured terrain has no lane lines and no map. Dust, mud, smoke, and low light defeat EO-only perception, and the terrain that matters most is remote, dangerous, and expensive to reach.

Stardust generates endless terrain, vegetation, and degraded-visibility conditions across EO, IR, LiDAR, and thermal, with point and pixel-level ground truth, so off-road perception and locomotion are tested before the field tests them.

Why real data falls short

// collecting off-road today

• ✕ Unstructured terrain has no lane lines and no map, every scene is different
• ✕ Dust, mud, and degraded visibility defeat EO-only perception
• ✕ The terrain that matters most is remote, dangerous, and expensive to reach

Photorealistic data, perfectly labeled

Unstructured terrain Generate endless variation in terrain, vegetation, and obstacles.

Degraded visibility Dust, mud, rain, smoke, and low light across EO, IR, and thermal.

LiDAR and thermal labels Registered multi-sensor data with pixel and point-level ground truth.

SENSOR COVERAGE

EO IR LiDAR THERMAL

What teams build with it

Terrain classification Drivable-surface and hazard segmentation off-road.

Obstacle detection Detect rocks, ditches, and vegetation in clutter.

Occlusion handling Perceive through dust, foliage, and partial views.

Learned locomotion Train policies for rough-terrain navigation and control.

Speaks your domain

The vocabulary, sensors, and benchmarks off-road teams actually use.

unstructured terrain traversability occlusion LiDAR thermal degraded visibility UGV learned locomotion semantic segmentation point cloud

Evaluate off-road policies on Manifold

Off-road autonomy is increasingly learned policy. Once perception is trained on Stardust, evaluate locomotion and navigation policies on Manifold.

Questions teams ask

How do you train off-road perception without field collection?

Generate endless unstructured terrain across dust, mud, smoke, and low light, with segmentation and point-level ground truth, then validate before you deploy.

Which sensors does off-road perception need?

EO alone is not enough off-road. Stardust generates registered EO, IR, LiDAR, and thermal so perception holds up in degraded visibility.

Can you train learned locomotion?

Yes. Train perception on Stardust, then evaluate locomotion and navigation policies on Manifold across terrain you would never risk in the field.

Test off-road autonomy before the terrain does

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