Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
library_name: lerobot
|
| 5 |
+
pipeline_tag: robotics
|
| 6 |
+
tags:
|
| 7 |
+
- vision-language-action
|
| 8 |
+
- imitation-learning
|
| 9 |
+
- lerobot
|
| 10 |
+
inference: false
|
| 11 |
+
license: gemma
|
| 12 |
+
base_model: lerobot/pi05_base
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# pi05_subtask base checkpoint
|
| 16 |
+
|
| 17 |
+
A derived copy of [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base) prepared for the
|
| 18 |
+
[`lerobot_policy_pi05_subtask`](https://github.com/) LeRobot plugin (π0.5 with System-1 subtask
|
| 19 |
+
generation, following [openpi_subtask](https://github.com/Ke-Wang1017/openpi_subtask)).
|
| 20 |
+
|
| 21 |
+
**The model weights are bit-identical to `lerobot/pi05_base`.** Only two metadata files differ:
|
| 22 |
+
|
| 23 |
+
- `config.json`: policy `type` is `pi05_subtask` instead of `pi05`.
|
| 24 |
+
- `policy_preprocessor.json`: pi05's `pi05_prepare_state_tokenizer_processor_step` +
|
| 25 |
+
`tokenizer_processor` steps are replaced by `pi05_subtask_tokenize_processor_step`, which builds
|
| 26 |
+
the two-segment prompt
|
| 27 |
+
|
| 28 |
+
```
|
| 29 |
+
context (no loss): "Task: {task}.; State: {256-bin ints}; Subtask: "
|
| 30 |
+
predicted (CE loss): "{subtask}.;\nAction: " + EOS
|
| 31 |
+
```
|
| 32 |
+
|
| 33 |
+
and emits a subtask region mask alongside the language tokens.
|
| 34 |
+
|
| 35 |
+
This exists because LeRobot instantiates processor pipelines from the checkpoint's
|
| 36 |
+
`policy_preprocessor.json` (by processor-step registry name), so fine-tuning the subtask policy
|
| 37 |
+
directly from `lerobot/pi05_base` would silently install plain pi05's tokenizer steps.
|
| 38 |
+
|
| 39 |
+
## Usage
|
| 40 |
+
|
| 41 |
+
Requires the `lerobot_policy_pi05_subtask` plugin installed (LeRobot auto-imports
|
| 42 |
+
`lerobot_policy_*` distributions). In a training config:
|
| 43 |
+
|
| 44 |
+
```json
|
| 45 |
+
"policy": {
|
| 46 |
+
"type": "pi05_subtask",
|
| 47 |
+
"pretrained_path": "dgrachev/pi05_subtask_base",
|
| 48 |
+
...
|
| 49 |
+
}
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
During training the subtask string is read from `complementary_data["subtask"]` when the dataset
|
| 53 |
+
provides one, and falls back to the task string otherwise.
|
| 54 |
+
|
| 55 |
+
Regenerate with `scripts/prepare_base_checkpoint.py` from the plugin repository.
|