---
library_name: pytorch
tags:
- robotwin
- vam
- wan2.2
- imitation-learning
- robotics
- mask-v3
- no-reference
---

# RoboTwin ICL V3 Mask No-Reference Checkpoint

This repository contains the 100k-step checkpoint for the RoboTwin ICL paired-v3
VAM training run with v3 masking and no reference video.

## Checkpoint

- File: `step-100000.safetensors`
- Dataset: `robotwin-icl-paired-v3`
- Target robot: `arx-x5`
- Task split: train20
- Mask variant: v3
- Reference setting: no reference video (`full_reference_video=false`)
- Action expert style: openwam
- Action/proprio space: eef
- Training budget: 100k steps, 2 nodes x 8 GPUs
- W&B run: https://wandb.ai/wuji-tech/vam_icl/runs/80vlexr0

The original local checkpoint path and run metadata are recorded in
`training_metadata.json`.