wam_attack

Motus + RoboTwin + white-box input attack (noise / patch) for robotic VLA policies.

Repository layout

wam_attack/
├── Motus/                 # Motus VLA model & training code
│   └── RoboTwin/          # RoboTwin simulation & evaluation
├── attacks/
│   ├── motus/             # Motus-specific attack train/eval scripts
│   └── fastwam/common/    # Shared noise/patch utilities

Setup

1. Clone this repo and create conda env (see Motus/requirements.txt, RoboTwin docs).

2. Download pretrained models into Motus/pretrained_models/:

   • Motus_robotwin2
   • Wan2.2-TI2V-5B
   • Qwen3-VL-2B-Instruct

3. Download RoboTwin assets (see Motus/RoboTwin/assets/_download.py).

4. Edit Motus/RoboTwin/policy/Motus/paths_config.yml with your local paths.

Attack workflow (Motus / RoboTwin)

conda activate robotwin
cd Motus/RoboTwin

# White-box noise training
GPU_ID=7 bash ../../attacks/motus/train.sh

# White-box patch training
GPU_ID=7 bash ../../attacks/motus/train_patch.sh

# Attack evaluation
GPU_ID=7 TASK_NAME=click_alarmclock bash ../../attacks/motus/eval.sh

# Clean baseline
bash policy/Motus/eval.sh

Notes

• Large weights, datasets, eval videos, and trained attack artifacts are not included in git.
• Patch/noise training optimizes universal perturbations against Motus action deviation (see attacks/motus/whitebox.py).