Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags: [openpi, pi0.5, robotics, vla, piper, cube-stacking]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# pi05_piper_stacking_v3 - step 2500 (early checkpoint)
|
| 7 |
+
|
| 8 |
+
pi0.5 (action_horizon 50) full fine-tune of pi05_base on
|
| 9 |
+
[axiboai/piper_stacking_v3](https://huggingface.co/datasets/axiboai/piper_stacking_v3)
|
| 10 |
+
(155 episodes cube stacking = v2 + ~44 correction episodes).
|
| 11 |
+
|
| 12 |
+
**FIRST checkpoint: step 2500 / 10000** (train loss ~0.009). Early smoke-test quality; finals follow.
|
| 13 |
+
Inference artifacts only (params/ + assets/norm_stats); no train_state/.
|
| 14 |
+
|
| 15 |
+
Prompt: `stack red cube on blue cube`
|
| 16 |
+
|
| 17 |
+
```python
|
| 18 |
+
import pathlib
|
| 19 |
+
from huggingface_hub import snapshot_download
|
| 20 |
+
from openpi.training import config as _config
|
| 21 |
+
from openpi.policies import policy_config
|
| 22 |
+
ckpt = pathlib.Path(snapshot_download("axiboai/pi05_piper_stacking_v3"))
|
| 23 |
+
policy = policy_config.create_trained_policy(_config.get_config("pi05_piper_stacking_v3"), ckpt)
|
| 24 |
+
```
|