Add model card for eval3_sanity checkpoints
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: lerobot
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- lerobot
|
| 6 |
+
- smolvla
|
| 7 |
+
- so101
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# eval3_sanity
|
| 11 |
+
|
| 12 |
+
All numbered SmolVLA overfit-20 checkpoints for Eval 3 sanity checking.
|
| 13 |
+
|
| 14 |
+
Dataset: `robot-learning-group47/eval3_overfit20`
|
| 15 |
+
|
| 16 |
+
Checkpoints:
|
| 17 |
+
|
| 18 |
+
- `checkpoints/000500`
|
| 19 |
+
- `checkpoints/001000`
|
| 20 |
+
- `checkpoints/001500`
|
| 21 |
+
- `checkpoints/002000`
|
| 22 |
+
- `checkpoints/002500`
|
| 23 |
+
- `checkpoints/003000`
|
| 24 |
+
|
| 25 |
+
The `003000/pretrained_model` checkpoint is also uploaded at repo root as the default loadable policy.
|
| 26 |
+
|
| 27 |
+
Key inference settings are stored in each checkpoint config:
|
| 28 |
+
|
| 29 |
+
- policy type: `smolvla`
|
| 30 |
+
- fps: `15`
|
| 31 |
+
- camera key: `observation.images.front`
|
| 32 |
+
- image size: `240x320`, resized internally with padding to `512x512`
|
| 33 |
+
- `chunk_size=50`
|
| 34 |
+
- `n_action_steps=50`
|
| 35 |
+
- `num_steps=10`
|
| 36 |
+
- `attention_mode=self_attn`
|
| 37 |
+
- `num_vlm_layers=8`
|
| 38 |
+
- `num_expert_layers=4`
|