---
license: apache-2.0
library_name: diffusers
base_model: robbyant/lingbot-va-base
datasets:
  - Wjjjh/libero_lingbot_va
tags:
  - robotics
  - vision-language-action
  - world-model
  - libero
  - lingbot-va
  - pytorch
---

# LingBot-VA LIBERO-Goal

This repository contains a community post-trained transformer checkpoint for
[LingBot-VA](https://github.com/robbyant/lingbot-va) on the LIBERO-Goal suite.
It is not an official Robbyant model release.

The checkpoint predicts robot actions jointly with future visual observations.
It is intended for research and evaluation in the LIBERO simulator.

## Checkpoint and result

| Repository path | Training step | LIBERO-Goal result |
|---|---:|---:|
| `transformer/` | 5,000 | 1,451 / 1,500 (96.73%, 3 runs) |

Each run covers all 10 LIBERO-Goal tasks with 50 episodes per task. The first
run used the runtime-default seed, which was not explicitly recorded; runs 2
and 3 used explicit seeds.

| Run | Seed | Successes / episodes | Success rate |
|---:|---|---:|---:|
| 1 | runtime default (not recorded) | 485 / 500 | 97.0% |
| 2 | `2026072402` | 486 / 500 | 97.2% |
| 3 | `2026072403` | 480 / 500 | 96.0% |
| **Pooled / mean** | 3 runs | **1,451 / 1,500** | **96.73%** |

The population standard deviation across the three run-level success rates is
0.52 percentage points.

For reference, the official LingBot-VA repository reports **97.2% ± 0.2%**
for LIBERO-Goal. This community checkpoint's three-run mean is 0.47 percentage
points lower. The two figures should not be treated as strictly identical
experiments unless every training and evaluation detail is matched.

## Model contents

This repository provides the post-trained **transformer only**. The VAE,
tokenizer, and text encoder must be taken from
[`robbyant/lingbot-va-base`](https://huggingface.co/robbyant/lingbot-va-base).

The checkpoint config is saved with:

```json
{"attn_mode": "flex"}
```

`flex` is the training setting. For inference or evaluation, copy the
checkpoint into a runtime model directory and change `attn_mode` to `torch` or
`flashattn`, following the official LingBot-VA instructions. Do not overwrite
the archived training checkpoint merely to switch inference backends.

## Download

```bash
hf download Wjjjh/lingbot-va-libero-goal \
  --include "transformer/*" \
  --local-dir lingbot-va-libero-goal-step5000
```

## Training details

| Setting | Value |
|---|---|
| Initialization | `robbyant/lingbot-va-base` |
| Dataset | `Wjjjh/libero_lingbot_va`, `libero_goal` subset |
| Dataset size | 500 episodes, 63,728 action rows |
| Optimizer | AdamW |
| Learning rate | `1e-5` |
| Betas | `(0.9, 0.95)` |
| Weight decay | `0.1` |
| Warmup | 10 optimizer steps |
| Precision | bfloat16 |
| Distributed strategy | FSDP |
| Hardware | 4 × NVIDIA H100 80 GB |
| Per-GPU batch size | 1 |
| Gradient accumulation | 8 |
| Effective global batch size | 32 |
| Total optimizer steps | 5,000 |
| Checkpoint interval | 200 steps |

Training followed the LingBot-VA LIBERO post-training pipeline, with a
suite-specific action normalization computed from all LIBERO-Goal training
actions.

## Action and evaluation configuration

The model uses the first seven channels of the 30-channel action layout:

```text
used_action_channel_ids = [0, 1, 2, 3, 4, 5, 6]
action_snr_shift = 0.05
```

The action quantiles used for both training and evaluation are:

```json
{
  "q01": [
    -0.8973214030265808,
    -0.7473214268684387,
    -0.9375,
    -0.15214285254478455,
    -0.16256785675883295,
    -0.14142857491970062,
    -1.0
  ],
  "q99": [
    0.9375,
    0.9133928418159485,
    0.9375,
    0.21214285492897034,
    0.2582142949104309,
    0.375,
    1.0
  ]
}
```

The reported evaluations used:

- LIBERO-Goal, 10 tasks
- 50 episodes per task (500 total)
- maximum 800 environment steps per episode
- 128 × 128 agent-view and eye-in-hand observations
- 4 parallel GPU shards
- one runtime-default seed and explicit seeds `2026072402` and `2026072403`
- LIBERO commit `8f1084e3132a39270c3a13ebe37270a43ece2a01`

## Limitations

- The first of the three evaluation runs used an unrecorded runtime-default
  seed; only runs 2 and 3 are exactly seed-reproducible from this card.
- Results are simulator- and configuration-dependent.
- This model has not been validated for safe real-world robot deployment.
- These files contain model weights only and do not contain optimizer state, so
  they are not a complete resumable training checkpoint.

## Acknowledgements and citation

This checkpoint is derived from the official
[LingBot-VA](https://github.com/robbyant/lingbot-va) base model and training
code. Please cite the original LingBot-VA work when using this model:

```bibtex
@article{lingbot-va2026,
  title={Causal World Modeling for Robot Control},
  author={Li, Lin and Zhang, Qihang and Luo, Yiming and Yang, Shuai and Wang, Ruilin and Han, Fei and Yu, Mingrui and Gao, Zelin and Xue, Nan and Zhu, Xing and Shen, Yujun and Xu, Yinghao},
  journal={arXiv preprint arXiv:2601.21998},
  year={2026}
}
```