---
license: apache-2.0
tags:
- lerobot
- pi0
- robotics
- moonbot
pipeline_tag: robotics
---

# moonbot pi0 — three blocks stack

pi0 fine-tuned from [`lerobot/pi0_base`](https://huggingface.co/lerobot/pi0_base)
on [`gdiazsrl/three_blocks_stack_sep22`](https://huggingface.co/datasets/gdiazsrl/three_blocks_stack_sep22)
(296 episodes, 519,684 frames, 9 block-stacking tasks).

## ⚠️ Checkpoint `020000` — training still in progress

This is an **intermediate checkpoint at step 20,000 of 20,000**, published
early so it can be evaluated while the run continues. It is not the final model.

## Training configuration

| | |
|---|---|
| base | `lerobot/pi0_base` |
| batch_size | 16 |
| steps (total) | 20,000 |
| dtype | bfloat16 |
| gradient_checkpointing | true |
| freeze_vision_encoder | false |
| train_expert_only | false |
| GPU | RTX PRO 6000 Blackwell (96 GB), ~39 GB used |

## Contents

`pretrained_model/` only — `model.safetensors` plus the pre/post-processor
configs. The optimizer state (~14 GB) is **not** included; this checkpoint is for
inference and evaluation, not for resuming training.

## Deployment note

Per Gustavo: open `config.json` in the downloaded checkpoint and **remove invalid
fields such as `pretrained_revision`** before loading with
`rosetta/policy_bridge_node.py`.

```bash
python src/rosetta/rosetta/policy_bridge_node.py --ros-args \
  -p contract_path:=<path>/moonbot1.yaml \
  -p policy_path:=<downloaded>/checkpoint-020000
```

The deployment contract must match the dataset: this model was trained on
**21-dim `observation.state`** (`joint_read` 8 + `tip_pos` 7 + `F_ee` 6).