Faster-WAM: Efficient Inference-Time Future Conditioning
for Robust World Action Models

Weiheng Zhao¹ · Haoyi Jiang¹ · Xin Shi² · Liu Liu³ · Zhizhong Su³ · Wei Sui² · Fan Huang⁴ · Xinggang Wang¹

Huazhong University of Science and Technology¹ · D-Robotics² · Horizon Robotics³ · Xiamen University⁴

The key insight behind Faster-WAM is that future representations are not merely an auxiliary training signal, but essential inference-time context for robust action prediction under distribution shifts. Guided by this principle, Faster-WAM computes future representations once and selectively reuses them during action denoising, reducing redundant video-action interaction. It achieves state-of-the-art in-distribution performance and robust OOD generalization across simulated and real-world manipulation, while substantially reducing inference latency.

---

Table of Contents

• Release Progress
• File Structure
• Environment Setup
• Model Preparation
• Dataset Download
• Training
• Released Checkpoints
• Evaluation
• Latency
• Acknowledgments
• Citation

Release Progress

• Training and inference code. [✔]
• LIBERO, LIBERO-Plus, and RoboTwin evaluation code. [✔]
• Model checkpoints. [✔]

File Structure

FasterWAM/
├── configs/
│   ├── data/                 # LIBERO and RoboTwin dataset configs
│   ├── model/                # FastWAM, JointWAM, and FasterWAM models
│   ├── task/                 # Benchmark-specific training configs
│   ├── sim_libero.yaml       # LIBERO evaluation defaults
│   ├── sim_libero_plus.yaml  # LIBERO-Plus evaluation defaults
│   └── sim_robotwin.yaml     # RoboTwin evaluation defaults
├── environments/             # Independent benchmark uv projects and locks
├── scripts/
│   ├── train.py
│   ├── train_zero1.sh
│   ├── preprocess_sparse_action_dit_backbone.py
│   ├── precompute_text_embeds.py
│   └── eval_fasterwam_*.sh
├── experiments/
│   ├── libero/               # LIBERO evaluation manager and worker
│   └── robotwin/             # RoboTwin evaluation manager and policy adapter
├── src/fasterwam/            # Core model, dataset, and training code
├── third_party/RoboTwin/     # RoboTwin evaluation integration
├── checkpoints/              # Wan components and model checkpoints
├── data/                     # Preprocessed training datasets
├── runs/                     # Training outputs
└── evaluate_results/         # Evaluation outputs

The final system is FasterWAM. The repository also keeps the FastWAM and JointWAM baselines for controlled comparisons.

Environment Setup

Run all commands below from the repository root.

Install uv first. The core training environment uses Python 3.10 and the PyTorch 2.7.1 CUDA 12.8 wheels locked in uv.lock:

bash scripts/setup/install_core.sh
source .venv/bin/activate

Model Preparation

FasterWAM uses Wan2.2-TI2V-5B. By default, missing components are downloaded from Hugging Face and stored under ./checkpoints. Set the directory explicitly before model preparation, training, or evaluation:

mkdir -p checkpoints
export DIFFSYNTH_MODEL_BASE_PATH="$(pwd)/checkpoints"

To use ModelScope instead, additionally set:

export DIFFSYNTH_DOWNLOAD_SOURCE=modelscope

FasterWAM ActionDiT initialization

Before training FasterWAM from scratch, generate its SparseActionDiT initialization from the Wan2.2 video DiT:

python scripts/preprocess_sparse_action_dit_backbone.py \
  --model-config configs/model/fasterwam.yaml \
  --output checkpoints/SparseActionDiT_cond_0_4_8_12_16_20_24_28_Wan22_alphascale_1024hdim.pt \
  --device cuda \
  --dtype bfloat16

Baseline ActionDiT initialization

FastWAM and JointWAM use the dense ActionDiT initialization:

python scripts/preprocess_action_dit_backbone.py \
  --model-config configs/model/fastwam.yaml \
  --output checkpoints/ActionDiT_linear_interp_Wan22_alphascale_1024hdim.pt \
  --device cuda \
  --dtype bfloat16

Dataset Download

LIBERO

FasterWAM uses the same preprocessed MuJoCo 3.3.2 LIBERO dataset as FastWAM:

• yuanty/LIBERO-fastwam

Download the four archives and extract them under data/libero_mujoco3.3.2:

mkdir -p data/libero_mujoco3.3.2

huggingface-cli download yuanty/LIBERO-fastwam \
  --repo-type dataset \
  --local-dir data/libero_mujoco3.3.2

cd data/libero_mujoco3.3.2
for f in *.tar.gz; do tar -xzf "$f"; done
cd ../..

The resulting layout must be:

data/libero_mujoco3.3.2/
├── libero_10_no_noops_lerobot/
├── libero_goal_no_noops_lerobot/
├── libero_object_no_noops_lerobot/
└── libero_spatial_no_noops_lerobot/

RoboTwin

The preprocessed RoboTwin dataset is available from:

• yuanty/robotwin2.0-fastwam

Download all split archives, concatenate them, and extract them as described in the FastWAM release:

mkdir -p data/robotwin2.0

huggingface-cli download yuanty/robotwin2.0-fastwam \
  --repo-type dataset \
  --local-dir data/robotwin2.0

cd data/robotwin2.0
cat robotwin2.0.tar.gz.part-* | tar -xzf -
cd ../..

The expected layout is:

data/robotwin2.0/
├── dataset_stats.json
└── robotwin2.0/
    ├── data/
    ├── meta/
    └── videos/

Training

Precompute instruction embeddings

Training reads cached T5 instruction embeddings. Generate them once after the dataset has been extracted:

# LIBERO
python scripts/precompute_text_embeds.py task=libero_fasterwam_2cam224_1e-4

# RoboTwin
python scripts/precompute_text_embeds.py task=robotwin_fasterwam_3cam_384_1e-4

For multi-GPU preprocessing:

torchrun --standalone --nproc_per_node=8 \
  scripts/precompute_text_embeds.py \
  task=libero_fasterwam_2cam224_1e-4

The caches are written to data/text_embeds_cache_fasterwam/libero and data/text_embeds_cache_fasterwam/robotwin.

Launch training

NPROC_PER_NODE=8 bash scripts/train_fasterwam_libero.sh

NPROC_PER_NODE=8 bash scripts/train_fasterwam_robotwin.sh

Both wrappers accept additional Hydra overrides. For example:

NPROC_PER_NODE=8 bash scripts/train_fasterwam_libero.sh \
  batch_size=8 \
  num_epochs=1 \
  wandb.enabled=true

Released Checkpoints

The released FasterWAM checkpoints and their corresponding dataset statistics are available on Hugging Face.

mkdir -p checkpoints/fasterwam_release

huggingface-cli download hustvl/FasterWAM \
  --local-dir checkpoints/fasterwam_release

After downloading, the checkpoint directory should have the following layout:

checkpoints/fasterwam_release/
├── libero/
│   ├── step_021700.pt
│   └── dataset_stats.json
└── robotwin/
    ├── step_029355.pt
    └── dataset_stats.json

Evaluation

LIBERO, LIBERO-Plus, and RoboTwin are managed in three separate uv environments to isolate their simulator dependencies. Run the corresponding setup command once before evaluation; each evaluation launcher automatically uses the matching environment.

LIBERO

bash scripts/setup/install_libero.sh

TASK_NAME=libero_fasterwam_2cam224_1e-4 \
CKPT_PATH=checkpoints/fasterwam_release/libero/step_021700.pt \
DATASET_STATS_PATH=checkpoints/fasterwam_release/libero/dataset_stats.json \
NUM_GPUS=8 \
bash scripts/eval_fasterwam_libero.sh

LIBERO-Plus

bash scripts/setup/install_libero_plus.sh

TASK_NAME=libero_fasterwam_2cam224_1e-4 \
CKPT_PATH=checkpoints/fasterwam_release/libero/step_021700.pt \
DATASET_STATS_PATH=checkpoints/fasterwam_release/libero/dataset_stats.json \
NUM_GPUS=8 \
bash scripts/eval_fasterwam_libero_plus.sh

RoboTwin

bash scripts/setup/install_robotwin.sh

TASK_NAME=robotwin_fasterwam_3cam_384_1e-4 \
CKPT_PATH=checkpoints/fasterwam_release/robotwin/step_029355.pt \
DATASET_STATS_PATH=checkpoints/fasterwam_release/robotwin/dataset_stats.json \
NUM_GPUS=8 \
bash scripts/eval_fasterwam_robotwin.sh

Latency

To measure the Inference Latency:

# All models
bash scripts/measure_latency.sh

# Selected models: jointwam, fastwam, fasterwam
bash scripts/measure_latency.sh --models jointwam fasterwam

Acknowledgments

Our codebase is built upon:

• FastWAM: https://github.com/yuantianyuan01/FastWAM
• Wan2.2: https://github.com/Wan-Video/Wan2.2
• LIBERO: https://github.com/Lifelong-Robot-Learning/LIBERO
• LIBERO-Plus: https://github.com/sylvestf/LIBERO-plus
• RoboTwin: https://github.com/RoboTwin-Platform/RoboTwin

We thank these teams for contributing their impressive code and models to the community.

Citation

If you find this repository helpful for your research, please consider citing our paper:

@article{zhao2026faster,
  title   = {Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models},
  author  = {Zhao, Weiheng and Jiang, Haoyi and Shi, Xin and Liu, Liu and Huang, Fan and Su, Zhizhong and Sui, Wei and Wang, Xinggang},
  journal = {arXiv preprint arXiv:2608.04404},
  year    = {2026}
}