Regolith: Lunar Hazard Segmentation from Synthetic Data
Randomized lunar scenes generated with NVIDIA Omniverse Replicator and OpenUSD on a DGX Spark, used to train a rock-hazard segmentation model on 100% synthetic data.
Regolith uses NVIDIA Omniverse Replicator, running headless inside Isaac Sim 6.0, with OpenUSD to generate randomized lunar scenes on our DGX Spark — 2,550 labeled images per build, with domain-randomized lighting, materials and camera. A SegFormer-B0 trained only on that synthetic data reaches 0.956 mIoU and 0.887 rock-IoU on the synthetic test set. We then ran it on 21 real NASA Apollo and Surveyor photographs as a qualitative check; those photos have no ground-truth labels, so the real-world results are inspection, not a score. The rover in the renders is a VIPER-style model we built ourselves. The project is written up as an 8-chapter series.