Zanyi Wang, Yuheng Lei, Dengyang Jiang, Ping Luo, Mengdi Wang, Zhixuan Liang, and Shilong Liu
Paper PDF · Quick start · Evaluation · Training · Citation
NowWAM adapts pretrained generative models to robot control through current-observation denoising, without a separate future-target branch. We provide model weights, data preparation, training and evaluation code.
📦 Klein weights · 🏋️ LIBERO & RoboCasa (100-shot) training · 🤖 LIBERO-Plus & RoboCasa evaluation
Use Python 3.11 on x86-64 Linux with an NVIDIA GPU and a working CUDA driver. On Ubuntu/Debian, install the system packages first:
sudo apt-get update
sudo apt-get install -y git build-essential python3.11-venv ffmpeg libgl1 libegl1 libglib2.0-0 libmagickwand-devAccept the model access terms for the FLUX.2 VAE if required, then install and authenticate:
git clone https://github.com/xmz111/NowWAM.git
cd NowWAM
python3.11 scripts/install.py Klein-LIBERO --dependencies-only
.venv/bin/hf auth login
python3.11 scripts/install.py Klein-LIBERO
source .venv/bin/activate
python -m eval Klein-LIBERO --limit 4The installer downloads the weights and assets, prepares the simulator, and
runs a one-episode test. Replace Klein-LIBERO with Klein-RoboCasa-100shot
to evaluate RoboCasa. Use all with the installer to prepare both checkpoints;
evaluate each by name.
Simulator environments are selected automatically. Do not reinstall during a run.
Full evaluation, with optional parallel lanes:
python -m eval Klein-LIBERO
# Or use two GPUs with two lanes each:
python -m eval Klein-LIBERO --gpus 0,1 --lanes 2Results are saved under outputs/<model>. Re-run the same command to resume.
Short tests use a separate output directory. Each parallel lane holds its own
policy; start with one lane if GPU memory is limited.
Keep the installed environment and assets unchanged while resuming.
For results from an older evaluator, use a new directory with --output outputs/new-run.
Run from the repository root. Use four GPUs with DDP (tested on 4 × H200,
141 GB each) and allow 500 GB for data, assets and checkpoints. Keep training in
its own environment; do not run pip install -e . inside it.
Accept any required base-model access terms
and the VAE terms above before downloading.
python3.11 -m venv .venv-train
source .venv-train/bin/activate
pip install -r requirements/train.txt
hf auth loginDownload the base models once; both training recipes use them:
Base-model download
hf download black-forest-labs/FLUX.2-klein-base-4B flux-2-klein-base-4b.safetensors \
--revision a3b4f4849157f664bdbc776fd7453c2783562f4d --local-dir assets/FLUX.2-klein-base-4B
hf download black-forest-labs/FLUX.2-dev ae.safetensors \
--revision 26afe3a78bb242c0a8bb181dcc8937bb16e5c66c --local-dir assets/FLUX.2-dev
hf download Qwen/Qwen3-4B --revision 1cfa9a7208912126459214e8b04321603b3df60c \
--include '*.json' '*.safetensors' 'merges.txt' 'vocab.json' 'tokenizer.model' --local-dir assets/Qwen3-4B
[ -d assets/sources/flux2/.git ] || git clone --no-checkout https://github.com/black-forest-labs/flux2.git assets/sources/flux2
git -C assets/sources/flux2 checkout --detach 50fe5162777813d869182b139e83b10743caef15
python -m training.prepareDownload the training data and train:
hf download yuanty/LIBERO-fastwam --repo-type dataset \
--revision ee018b997c430bb12b5bf3c892d744798c5a2f91 \
--include 'libero_*_no_noops_lerobot.tar.gz' --local-dir data/archives
for suite in spatial object goal 10; do
tar -xzf "data/archives/libero_${suite}_no_noops_lerobot.tar.gz" -C data
done
OMP_NUM_THREADS=4 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc-per-node=4 -m trainingPrepare the RoboCasa simulator, then download and process the training data
from .venv-train:
python3.11 scripts/install.py Klein-RoboCasa-100shot
python -m training.robocasa_data
OMP_NUM_THREADS=4 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc-per-node=4 -m training.robocasaPreparation downloads the GR1 demonstrations,
uses episodes 0–99 per task, and prepares observations, actions, statistics and
text features under data/robocasa.
Both recipes use global batch 64 and 42,210 updates. Repeat the training command
to resume, or append a new output directory for a new run. EMA policies are
exported under outputs/<benchmark>/exports/. After evaluation setup, run:
.venv/bin/python -m eval outputs/libero/exports/step_032830 --gpus 0,1 --lanes 2
.venv/bin/python -m eval outputs/robocasa/exports/step_037520 --gpus 0,1 --lanes 2@article{wang2026nowwam,
title = {Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control},
author = {Wang, Zanyi and Lei, Yuheng and Jiang, Dengyang and Luo, Ping and Wang, Mengdi and Liang, Zhixuan and Liu, Shilong},
journal = {arXiv preprint arXiv:2609.28339},
year = {2026}
}We thank FastWAM and ImageWAM for their open-source work, and TRC for compute support.