File size: 2,975 Bytes
1b0f2f6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 | ---
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/*",
)
```
|