Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- lerobot
|
| 5 |
+
- pi0
|
| 6 |
+
- robotics
|
| 7 |
+
- moonbot
|
| 8 |
+
pipeline_tag: robotics
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# moonbot pi0 — three blocks stack
|
| 12 |
+
|
| 13 |
+
pi0 fine-tuned from [`lerobot/pi0_base`](https://huggingface.co/lerobot/pi0_base)
|
| 14 |
+
on [`gdiazsrl/three_blocks_stack_sep22`](https://huggingface.co/datasets/gdiazsrl/three_blocks_stack_sep22)
|
| 15 |
+
(296 episodes, 519,684 frames, 9 block-stacking tasks).
|
| 16 |
+
|
| 17 |
+
## ⚠️ Checkpoint `005000` — training still in progress
|
| 18 |
+
|
| 19 |
+
This is an **intermediate checkpoint at step 5,000 of 20,000**, published
|
| 20 |
+
early so it can be evaluated while the run continues. It is not the final model.
|
| 21 |
+
|
| 22 |
+
## Training configuration
|
| 23 |
+
|
| 24 |
+
| | |
|
| 25 |
+
|---|---|
|
| 26 |
+
| base | `lerobot/pi0_base` |
|
| 27 |
+
| batch_size | 16 |
|
| 28 |
+
| steps (total) | 20,000 |
|
| 29 |
+
| dtype | bfloat16 |
|
| 30 |
+
| gradient_checkpointing | true |
|
| 31 |
+
| freeze_vision_encoder | false |
|
| 32 |
+
| train_expert_only | false |
|
| 33 |
+
| GPU | RTX PRO 6000 Blackwell (96 GB), ~39 GB used |
|
| 34 |
+
|
| 35 |
+
## Contents
|
| 36 |
+
|
| 37 |
+
`pretrained_model/` only — `model.safetensors` plus the pre/post-processor
|
| 38 |
+
configs. The optimizer state (~14 GB) is **not** included; this checkpoint is for
|
| 39 |
+
inference and evaluation, not for resuming training.
|
| 40 |
+
|
| 41 |
+
## Deployment note
|
| 42 |
+
|
| 43 |
+
Per Gustavo: open `config.json` in the downloaded checkpoint and **remove invalid
|
| 44 |
+
fields such as `pretrained_revision`** before loading with
|
| 45 |
+
`rosetta/policy_bridge_node.py`.
|
| 46 |
+
|
| 47 |
+
```bash
|
| 48 |
+
python src/rosetta/rosetta/policy_bridge_node.py --ros-args \
|
| 49 |
+
-p contract_path:=<path>/moonbot1.yaml \
|
| 50 |
+
-p policy_path:=<downloaded>/checkpoint-005000
|
| 51 |
+
```
|
| 52 |
+
|
| 53 |
+
The deployment contract must match the dataset: this model was trained on
|
| 54 |
+
**21-dim `observation.state`** (`joint_read` 8 + `tip_pos` 7 + `F_ee` 6).
|