---
library_name: pytorch
tags:
  - robotics
  - world-model
  - vla
  - fastwam
  - lingbot-va
datasets:
  - jokeru/pick_right_purple_box_and_place_it_in_the_middle
---

# FastWAM and LingBot-VA Purple Box Checkpoints

Public release of two checkpoints fine-tuned on
`jokeru/pick_right_purple_box_and_place_it_in_the_middle`.

## Contents

| Path | Model | Training point | Notes |
| --- | --- | ---: | --- |
| `fastwam/step_000500.pt` | memory FastWAM | step 500 | Full model checkpoint used by the local FastWAM runtime |
| `lingbot-va/transformer/` | old `Robbyant/lingbot-va` | step 250 | Fine-tuned transformer weights and transformer config |

The LingBot-VA optimizer/FSDP state (`training_state.pt`, about 40.5 GB) is
intentionally excluded. The uploaded transformer is sufficient for inference
and as a fine-tuning initialization, but still requires the upstream
LingBot-VA code and base Wan/VAE/text-encoder components.

## Data and split

- Source dataset: `jokeru/pick_right_purple_box_and_place_it_in_the_middle`
- Training episodes: 18
- Held-out episodes: 4 and 15
- Task: pick the right purple box and place it in the middle

## FastWAM

- Architecture: project memory FastWAM variant
- Input views: `global_primary`, `left_wrist`, `right_wrist`
- Video size: 384 x 320
- Action/proprio schema: canonical 80D with validity masks
- Context length: 128
- Memory history: long 8, mid 2, short 1
- Optimization: 8 GPUs, batch 8/GPU, global batch 64, bf16, 500 steps
- Initialization: official RoboTwin FastWAM MoT checkpoint plus the Helios
  memory patchers from memory checkpoint step 35000

The detailed transfer metadata is stored in
`fastwam/initialization_report.json`. Paths in that report have been reduced
to stable checkpoint identifiers.

## LingBot-VA

- Architecture: old `Robbyant/lingbot-va`
- Optimization: 8 GPUs, batch 24/GPU, global batch 192, 250 steps
- Fixed latent window: 16 frames
- Action alignment: 16 source action steps per latent frame
- Training episodes: 18; held-out episodes: 4 and 15
- Input views: `right_eye`, `left_wrist`, `right_wrist`
- Compact action: right-arm joints 7D plus right gripper 1D, mapped into the
  model's 30D action schema

## Held-out open-loop results

These numbers are diagnostics from eight generated cases, not closed-loop
robot success rates.

| Model | Generated frames | Video PSNR | Video MAE | Action MAE |
| --- | ---: | ---: | ---: | ---: |
| FastWAM | 101 | 16.727 | 0.0776 | 0.4830 |
| LingBot-VA | 141 | 16.540 | 0.0913 | 0.1146 |

## Loading

Download the FastWAM checkpoint:

```python
from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="PencilHu/FastWAM-LingBotVA-PurpleBox",
    filename="fastwam/step_000500.pt",
)
```

Download the LingBot-VA transformer snapshot:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="PencilHu/FastWAM-LingBotVA-PurpleBox",
    allow_patterns="lingbot-va/transformer/*",
)
```