Reports — curriculum TOC
Read in order 00 → 07. Each chapter is self-contained but assumes the previous one. The series turns on a surprising sim-to-real failure (the photoreal-rocks build, v2) and then resolves it — the failure’s diagnosis becomes the v3 fix. Read for the arc, not a single trophy.
| # | Chapter | What you build |
|---|---|---|
| 00 | The primer | Background: semantic segmentation, synthetic data, sim-to-real gap, domain randomization, and the DGX Spark |
| 01 | The lunar stage | A procedural OpenUSD scene: regolith heightfield, realistic noise-displaced basalt rocks, sun, dome, rover camera; first pixel-accurate segmentation mask |
| 02 | Domain randomization | Replicator pipeline; 2,550 labeled frames across three splits (DR, no-DR, unseen-domain test) with randomized lighting, materials, and geometry |
| 03 | Training | SegFormer-B0 fine-tuned on synthetic data; honest size-matched ablation isolating domain randomization from dataset size |
| 04 | Sim-to-real | Transfer evaluation on real Apollo lunar photographs — the photoreal-rocks (v2) model floods ~83% of real regolith as rock |
| 05 | The render | Cinematic 1920×1080 RTX flythrough with live dr_1500 hazard overlay — clean, because in-distribution |
| 06 | Rock fidelity | The v1→v2 evolution and the failure: higher rock fidelity raised synthetic scores but worsened real transfer |
| 07 | Realistic ground | v3: the fix — fidelity moved to the ground (cratered dark displaced regolith). Synth rock-IoU 0.887, real flood 35.7% (v2 was ~83%); VIPER render |
Heavy artifacts (datasets, checkpoints, full-res render) live on the DGX Spark, not in git. Each report’s ## Reproduce section has the copy-pasteable ssh spark "docker exec …" commands.