LIQIIIII/badwam-robotwin-idm-wam

This repository contains the IDM WAM checkpoint used for RobotWin experiments of:

BadWAM: World-Action Desynchronization Attacks against World-Action Models

Files

File	Description
model.pt	Final IDM WAM checkpoint trained on RobotWin with the FastWAM robotwin_idm_3cam_384_1e-4 configuration.
dataset_stats.json	RobotWin dataset statistics required for action normalization and denormalization.
model_index.json	Machine-readable metadata for this checkpoint.
config.json	Loading manifest identifying the checkpoint, normalization statistics, and model variant.

3cam_384 denotes the three-camera RobotWin setup with 384-pixel visual inputs.

This upload replaces the previous in-progress checkpoint at step 132500 with the final checkpoint at step 234830.

Related checkpoints

• Official FastWAM action-only / unconditioned RobotWin checkpoint: yuanty/fastwam
  • robotwin_uncond_3cam_384.pt
• BadWAM RobotWin Joint WAM checkpoint: LIQIIIII/badwam-robotwin-joint-wam
• BadWAM LIBERO Joint WAM checkpoint: LIQIIIII/badwam-libero-joint-wam
• BadWAM LIBERO IDM WAM checkpoint: LIQIIIII/badwam-libero-idm-wam

Download the complete model bundle

Request access on this model page and wait for approval, then sign in with hf auth login. Download the complete repository so that the checkpoint, normalization statistics, and loading manifest stay together:

hf download LIQIIIII/badwam-robotwin-idm-wam --local-dir ./checkpoints/badwam-robotwin-idm-wam

The root config.json is a loading manifest for the native BadWAM/FastWAM code. It identifies the checkpoint and normalization-statistics files; the existing model_index.json remains available for compatibility.

Alternatively, use Python to download the bundle and resolve its files from the manifest:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

folder = Path(snapshot_download(
    repo_id="LIQIIIII/badwam-robotwin-idm-wam",
    local_dir="./checkpoints/badwam-robotwin-idm-wam",
))
config = json.loads((folder / "config.json").read_text())
checkpoint_path = folder / config["checkpoint_file"]
dataset_stats_path = folder / config["dataset_stats_file"]

Use these paths with the evaluation code, or arrange them in the local layout below.

Suggested local layout

For BadWAM/FastWAM-style evaluation, place or symlink the files as:

local_checkpoints/trained_robotwin/
  robotwin_idm_3cam_384.pt
  robotwin_idm_3cam_384_dataset_stats.json

Example:

mkdir -p local_checkpoints/trained_robotwin
ln -s /path/to/model.pt local_checkpoints/trained_robotwin/robotwin_idm_3cam_384.pt
ln -s /path/to/dataset_stats.json local_checkpoints/trained_robotwin/robotwin_idm_3cam_384_dataset_stats.json

Citation

@article{li2026badwam,
  title={BadWAM: When World-Action Models Dream Right but Act Wrong},
  author={Li, Qi and Yang, Xingyi and Wang, Xinchao},
  journal={arXiv preprint arXiv:2607.15207},
  year={2026}
}