---
library_name: diffusers
pipeline_tag: robotics
license: other
base_model: Miical/pi05-base
datasets:
  - lerobot/libero_10_image
tags:
  - robotics
  - vision-language-action
  - pi0.5
  - libero
  - supervised-fine-tuning
  - reinforcement-learning
---

# PI0.5 LIBERO-10 SFT — Step 200

This is a deliberately undertrained PI0.5 checkpoint fine-tuned on the complete
official LIBERO-10 dataset. It is intended as a reproducible starting policy
for measuring how much downstream reinforcement learning, particularly ReCap,
improves long-horizon task success rather than as a fully trained SFT release.

The checkpoint is an upstream-native, merged PI0.5 policy exported by
[verl-vla](https://github.com/verl-vla/verl-vla). It does not require a
verl-vla checkpoint wrapper for inference.

## Lineage

- Base model: [`Miical/pi05-base`](https://huggingface.co/Miical/pi05-base)
- Training dataset: [`lerobot/libero_10_image`](https://huggingface.co/datasets/lerobot/libero_10_image)
- Benchmark: LIBERO-10, all 10 tasks
- Checkpoint: optimizer step 200

## Training configuration

| Setting | Value |
| --- | --- |
| Training method | Supervised fine-tuning (SFT) |
| Optimizer steps | 200 |
| Dataset size | 101,469 frame-level samples, 379 episodes |
| Samples consumed | 51,200 |
| Effective epochs | approximately 0.50 |
| Global / mini-batch size | 256 |
| Micro-batch size | 16 |
| GPUs | 8 x NVIDIA H20 |
| Distributed strategy | FSDP2 |
| Model dtype | bfloat16 |
| Learning rate | 1e-4 |
| Weight decay | 1e-5 |
| LR warmup ratio | 0.05 |
| Action horizon | 10 |
| DataLoader workers | 8 |
| Seed | 42 |

One complete pass over this dataset is 396 optimizer steps with the configured
batch size. This checkpoint was saved at step 200, about halfway through the
first epoch. It therefore provides a moderately capable but intentionally
non-saturated policy for subsequent reinforcement-learning experiments.

## LIBERO-10 evaluation

Evaluation used all 10 LIBERO-10 tasks with 10 trials per task, a maximum of
520 environment steps per trial, and 52 policy interactions with a 10-action
chunk.

| Task | Instruction | Successes | Success rate |
| ---: | --- | ---: | ---: |
| 0 | Put both the alphabet soup and the tomato sauce in the basket | 4 / 10 | 40% |
| 1 | Put both the cream cheese box and the butter in the basket | 6 / 10 | 60% |
| 2 | Turn on the stove and put the moka pot on it | 4 / 10 | 40% |
| 3 | Put the black bowl in the bottom drawer of the cabinet and close it | 5 / 10 | 50% |
| 4 | Put the white mug on the left plate and put the yellow and white mug on the right plate | 3 / 10 | 30% |
| 5 | Pick up the book and place it in the back compartment of the caddy | 8 / 10 | 80% |
| 6 | Put the white mug on the plate and put the chocolate pudding to the right of the plate | 6 / 10 | 60% |
| 7 | Put both the alphabet soup and the cream cheese box in the basket | 5 / 10 | 50% |
| 8 | Put both moka pots on the stove | 0 / 10 | 0% |
| 9 | Put the yellow and white mug in the microwave and close it | 2 / 10 | 20% |
| **Overall** | **All tasks** | **43 / 100** | **43%** |

These results are from one complete 100-trajectory benchmark run. Per-task
estimates use only 10 trials and should not be interpreted as precise
population rates.

## Intended use

This model is intended as an incompletely trained initial policy for validating
whether ReCap or another reinforcement-learning method improves aggregate and
per-task LIBERO-10 success. In particular, task 8 has no successful trajectory
in this evaluation, while the other tasks provide varying amounts of initial
capability. For a high-performing SFT baseline or deployment candidate, use a
later and more thoroughly evaluated checkpoint instead.

## Loading

```python
from giga_models import PI0Policy

policy = PI0Policy.from_pretrained(
    "Miical/pi05-libero-10-sft-step-200"
)
```

With verl-vla, set the model path directly:

```yaml
cluster:
  actor_rollout_ref:
    model:
      path: Miical/pi05-libero-10-sft-step-200
      adapter:
        embodiment: libero
```

## License and attribution

This fine-tuned checkpoint inherits the usage and redistribution constraints of
the original PI0.5 checkpoint and its underlying components. The PyTorch
conversion and verl-vla integration code have their own respective licenses;
consult the base model card and upstream projects before redistribution or
commercial use.