Synthetic data for space autonomy

Rendezvous, pose estimation, and hazard detection for orbits and bodies where real training data does not exist.

Train for missions that have never flown.

If you cannot collect it, simulate it

There is no real dataset for a spacecraft that has not flown, an orbit no camera has seen, or a landing site no probe has touched. Lighting is extreme and unforgiving, with hard shadows and no atmosphere, and missions are one-shot, so perception has to work the first time.

Stardust renders physically accurate space scenes with 6-DoF ground truth, so teams can train and validate rendezvous, pose estimation, and hazard detection against conditions that exist nowhere else.

Why real data falls short

// training for space today

• ✕ There is no real dataset for a novel spacecraft, orbit, or landing site
• ✕ Lighting is extreme and unforgiving, with hard shadows and no atmosphere
• ✕ Missions are one-shot, so perception has to work the first time

Photorealistic data, perfectly labeled

Physically-accurate lighting Model harsh, atmosphere-free illumination, eclipse, and extreme dynamic range.

Novel objects and orbits Generate spacecraft, debris, and terrain that have never been imaged.

Ground truth in 6-DoF Pose, range, and segmentation labels for rendezvous and landing.

What teams build with it

Rendezvous and proximity ops Relative navigation and docking perception.

Pose estimation 6-DoF pose of resident space objects and targets.

Terrain and landing hazard Detect hazards for descent and landing.

RSO tracking Detect and track resident space objects and debris.

Speaks your domain

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

Working with the organizations operating in the hardest environment there is.

Questions teams ask

How do you train perception for a spacecraft that has not flown?

Render the spacecraft, orbit, and lighting in simulation with 6-DoF ground truth, and train pose and detection models against scenes that have no real equivalent.

Why is space lighting hard for perception?

There is no atmosphere to soften it, so you get hard shadows, extreme dynamic range, and specular glints. Stardust models these physically so models do not break on orbit.

Can you generate data for rendezvous and docking?

Yes. Relative pose, range, and segmentation labels for proximity operations, across approach geometries and lighting.