Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# eval2_90
|
| 2 |
+
|
| 3 |
+
SmolVLA policy for Eval 2.
|
| 4 |
+
|
| 5 |
+
Dataset:
|
| 6 |
+
- `robot-learning-group47/eval2_90_chengming_promptaug_v1`
|
| 7 |
+
|
| 8 |
+
Training setup:
|
| 9 |
+
- Base policy: `lerobot/smolvla_base`
|
| 10 |
+
- One real camera: `observation.images.front`
|
| 11 |
+
- Camera mapping: `observation.images.front` → `observation.images.camera1`
|
| 12 |
+
- `empty_cameras=2`
|
| 13 |
+
- Dynamic prompt sampling: random prompt sampled within the same layout-target bucket during training
|
| 14 |
+
- `n_obs_steps=1`
|
| 15 |
+
- `chunk_size=50`
|
| 16 |
+
- `n_action_steps=50`
|
| 17 |
+
- Batch size: 64
|
| 18 |
+
- Steps: 30,000
|
| 19 |
+
- Save frequency: 2,500
|
| 20 |
+
- `train_expert_only=true`
|
| 21 |
+
- `freeze_vision_encoder=true`
|
| 22 |
+
- `num_vlm_layers=16`
|
| 23 |
+
- Learning rate: `1e-4`
|
| 24 |
+
- Weight decay: `1e-10`
|
| 25 |
+
|
| 26 |
+
The repo root contains the latest checkpoint for direct loading.
|
| 27 |
+
|
| 28 |
+
All intermediate checkpoints are stored under:
|
| 29 |
+
|
| 30 |
+
`checkpoints/`
|