---
tags:
- robotics
- video-generation
- vision-language-action
- fruit-picking
library_name: pytorch
---

# Fruit-picking Flashwam

Public archival bundle for the fruit-picking model trained by Amin.

## Status

This is a **partial checkpoint: step 12,600 of 18,900 (epoch 20 of 30)**. The job terminated when its shared `/dev/shm`
video staging disappeared; it did not finish the planned 30 epochs. The model
has not been evaluated on a robot, and the physical meaning of the `-1/+1`
gripper polarity still needs confirmation.

Architecture: FlashWAM M1 fused-KV, fixed-RoPE, 1-layer action expert.

Checkpoints follow the same layout as the existing WAM repositories:

- `checkpoints/weights/step_003150.pt` — SHA-256 `2dc3d4ce58a592944975fc6aa2214f3f22cf724b08fa3815588d4d83319e31b0`
- `checkpoints/weights/step_006300.pt` — SHA-256 `660268a99a6e298ad5e3948f4784e7d8be6b4e0ad0517d9404b52c1dc7797145`
- `checkpoints/weights/step_009450.pt` — SHA-256 `dc0497a96a8c2c7bcf072d00e700302707f260bf6be0c7fa77567520f101ca01`
- `checkpoints/weights/step_012600.pt` — SHA-256 `96c9587141286f2bef8f7c38bc31816e5f9ce689f3c66500b5fdf1d6f3b83a06`

## Conditioning

Exact task text:

> Lift the lid, put it aside, and pick the black plum.

`conditioning/text_embedding.pt` is the exact cached T5 embedding consumed
during training. It was generated with the Wan text stack, context length 128,
using `Wan-AI/Wan2.1-T2V-1.3B` as the tokenizer model reference. The resolved
training config sets `load_text_encoder: false`, so this cached tensor is part
of the required inference bundle.

## Input processing and normalization

- Two 256x256 RGB cameras (`agentview`, then `wrist`).
- Each camera is converted to a tensor and resized to 224x224.
- Cameras are concatenated horizontally to 224x448.
- Horizon: 33 observations; 32 action transitions at 10 Hz.
- Original 15-D state was converted to 8-D:
  `eef_xyz(3) + quat-to-axis-angle(3) + [gripper_width/2, -gripper_width/2]`.
- Action is 7-D: delta XYZ, delta rotation XYZ, and gripper.
- Delta/padding mask is `[true, true, true, true, true, true, false]`; the
  gripper channel is absolute rather than delta.
- `dataset_stats.json` contains the exact min/max normalization statistics
  used by this run.

## Attention masks

The resolved model uses:

- `video_attention_mask_mode: first_frame_causal`: first-frame queries cannot
  attend to later video frames; later-frame queries can attend to all video
  tokens.
- `action_group_causal_mask_mode: group_diagonal`: each video temporal group
  attends only to the corresponding action-token group.
- Text cross-attention is enabled for the action expert.

The exact implementations and preprocessing classes are included under
`training_code/`; the resolved config is `config.yaml`.
No license is asserted here for the bundled upstream code; its original terms
continue to apply.

## Base components

This weights-only checkpoint is not standalone. It references
`Wan-AI/Wan2.2-TI2V-5B` and requires the matching Wan VAE plus the included
FastWAM code/configuration. PyTorch `.pt` files may contain pickled objects;
load only in a trusted environment.