Upload FunctionalPourMug_v1/shadow_right-best-100/README.md with huggingface_hub
Browse files
FunctionalPourMug_v1/shadow_right-best-100/README.md
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# FunctionalPourMug_v1/shadow_right-best-100
|
| 2 |
+
|
| 3 |
+
Image Diffusion Policy (official Chi et al. recipe, one ResNet18 per camera) on **Dexverse-FunctionalPourMug-v1**, hand **shadow_right**,
|
| 4 |
+
author v1 demos (48 train / 2 val), EMA weights at epoch **100** / 300.
|
| 5 |
+
|
| 6 |
+
- cameras: ('rgb_image', 'wrist_rgb_image') (stored 128, crop 116), state groups ('proprio', 'goal')
|
| 7 |
+
- obs / action horizon 2 / 16, U-Net chi (512, 1024, 2048)
|
| 8 |
+
- training run: `20261016-091041_vast-isaac_dp-img-v1-functionalpourmug` (hscho_share experiments)
|
| 9 |
+
- inference: DDPM 100 steps, replan every 8 steps (`scripts/diffusion/eval_online.py --observation_preset rgb`)
|
| 10 |
+
- code: DexVerse branch `dp/image-v1`
|
| 11 |
+
- role: best (highest success among the epoch 50-250 snapshots in 10 selection rollouts (seed 2000), ties to the later epoch)
|
| 12 |
+
|
| 13 |
+
- moved from the former repo `Dexverse-CSI/dp_image-shadow_right-FunctionalPourMug_v1-100-best`
|