Add Core4 MLP G2 25k artifact core4_mlp_current_state_g2_step25000/README.md

core4_mlp_current_state_g2_step25000/README.md

@@ -0,0 +1,17 @@
+# Core4 current-state MLP G2 @ 25k
+
+DexFlowMLP Stage-2 (action-only L1, no aux). Human Stage-1 init is
+`stage1_step02000.ckpt` (Future Articulated HandFlow + Current Object State).
+Robot data: four Core4 tasks × 10 action-only episodes. Same-set open-loop,
+not held-out SR.
+
+| Task folder | Open-loop nRMSE41 | body RMSE (rad) | hand MAE | chunk0 body / hand |
+|---|---:|---:|---:|---:|
+| `bottle_upright_right_crop` | 0.050 | 0.0050 | 2.74 | 0.0059 / 3.15 |
+| `pill_bottle_right_crop` | 0.067 | 0.0059 | 4.16 | 0.0063 / 4.44 |
+| `open_box_lid_both_crop` | 0.054 | 0.0089 | 3.01 | 0.0118 / 3.40 |
+| `realsense_box_into_box_right_crop` | 0.063 | 0.0062 | 3.63 | 0.0073 / 3.86 |
+
+Each folder: `stage2_step25000.ckpt` + same-run `norm_stats.json`.
+Load with `DexFlowMLPPolicy.from_checkpoint(..., stage="deploy")`.
+Publish the full 30-step chunk; no temporal ensemble.