---
license: mit
library_name: lerobot
tags:
  - robotics
  - vision-language-action
  - libero
  - smolvla
  - vgas
---

# VGAS and VGAS+ — 5-Shot LIBERO Checkpoints

This repository contains the checkpoints used to evaluate
[VGAS](https://arxiv.org/abs/2602.07399) and VGAS+ on the 5-shot LIBERO benchmark.
The implementation is available in the
[VGAS code repository](https://github.com/Jyugo-15/VGAS).

## Repository layout

```text
smolvla/5_SHOT/pretrained_model/       # Shared 5-shot SmolVLA policy
{suite}/vgas_critic/last.ckpt          # VGAS inference-time critic
{suite}/vgas_plus/pretrained_model/    # Distilled VGAS+ policy
```

Here, `{suite}` is one of `goal`, `object`, `spatial`, or `long` (`long` corresponds
to `libero_10`). VGAS uses the shared SmolVLA policy together with the suite-specific
critic for Best-of-N selection. The SFT policy is frozen while training the VGAS critic.
VGAS+ directly executes the distilled policy and does not require a critic or
inference-time reranking.

## VGAS+ checkpoints

| Directory | LIBERO suite |
|---|---|
| `goal/vgas_plus/pretrained_model` | `libero_goal` |
| `object/vgas_plus/pretrained_model` | `libero_object` |
| `spatial/vgas_plus/pretrained_model` | `libero_spatial` |
| `long/vgas_plus/pretrained_model` | `libero_10` |

Download one policy with `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="SemyonXu616/VGAS-5-shot",
    allow_patterns=["goal/vgas_plus/pretrained_model/*"],
    local_dir="checkpoints/VGAS-5-shot",
)
```

The downloaded policy directory can be passed directly as `POLICY_PATH` to the
VGAS+ evaluation scripts in the code repository.

## Citation

```bibtex
@article{xu2026vgas,
  title   = {VGAS: Value-Guided Action-Chunk Selection for Few-Shot Vision-Language-Action Adaptation},
  author  = {Xu, Changhua and Yu, En and Xuan, Junyu and Lu, Jie},
  journal = {arXiv preprint arXiv:2602.07399},
  year    = {2026}
}
```

The VGAS+ citation will be added when its preprint is available.