Analysis of internal representations learned by robot foundation models (D-LAPA pipeline).
Each research question maps to one numbered sub-folder, ordered by execution date.
| Folder | Research Question | Dataset | Date |
|---|---|---|---|
| 01_geometric_motion_probe | Which robot model features correlate most strongly with geometric motion (translation magnitude & direction)? | LIBERO-100 test + CALVIN val + LIBERO-10 val | 2026-06-24 → 2026-06-27 |
| 02_latent_space_umap | Does the latent space of Model 4 form a smooth gradient by motion magnitude? (motion-aware structure) | LIBERO-10 val (8 k subsample) | 2026-06-27 |
| 03_information_theory | Is the combination of RGB + Depth synergistic or redundant? Mutual Information analysis. | LIBERO-10 val (20 k subsample) | 2026-06-27 |
| 04_model_depth_analysis | Why does Model 4 outperform Model 2? Zero-out ablation, noise sensitivity, cluster separation. | LIBERO-10 val (15 k subsample) | 2026-06-27 |
<index>_<short_descriptive_name>