Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- lerobot
|
| 6 |
+
- smolvla
|
| 7 |
+
- bimanual
|
| 8 |
+
- vlm-frozen
|
| 9 |
+
datasets:
|
| 10 |
+
- Bcryan/BIMAN_PICK_AND_PLACE2
|
| 11 |
+
library_name: lerobot
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# SmolVLA Fine-tuned (VLM Frozen)
|
| 15 |
+
|
| 16 |
+
Fine-tuned SmolVLA model with frozen Vision-Language Model on BIMAN_PICK_AND_PLACE2 dataset.
|
| 17 |
+
|
| 18 |
+
## Model Configuration
|
| 19 |
+
- **freeze_vision_encoder**: True
|
| 20 |
+
- **train_expert_only**: True
|
| 21 |
+
- **train_state_proj**: True
|
| 22 |
+
|
| 23 |
+
## Training Details
|
| 24 |
+
- Base model: lerobot/smolvla_base
|
| 25 |
+
- Dataset: Bcryan/BIMAN_PICK_AND_PLACE2
|
| 26 |
+
- Training steps: 70,000
|
| 27 |
+
- Batch size: 16
|
| 28 |
+
- VLM: Frozen (only action expert trained)
|
| 29 |
+
|
| 30 |
+
## Usage
|
| 31 |
+
```python
|
| 32 |
+
from lerobot.common.policies.smolvla import SmolVLAPolicy
|
| 33 |
+
policy = SmolVLAPolicy.from_pretrained("Autobrik/smolvla-finetunned-vlm-off")
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
## Training Command
|
| 37 |
+
```bash
|
| 38 |
+
python lerobot/scripts/train.py \
|
| 39 |
+
--policy.path=lerobot/smolvla_base \
|
| 40 |
+
--dataset.repo_id=Bcryan/BIMAN_PICK_AND_PLACE2 \
|
| 41 |
+
--batch_size=16 \
|
| 42 |
+
--steps=70000
|
| 43 |
+
```
|