World Guidance: World Modeling in Condition Space for Action Generation

This is an unofficial implemention of WoG based on open source framework and datasets.

Thanks to OpenVLA and CogACT for their awesome codebase.

Installation

conda create -n wog python=3.10

Next, clone our repo and install the required packages:

git clone https://github.com/Selen-Suyue/WoG
cd WoG
pip install -e .
make setup

Flash Attention is needed for training. You can simply run:

pip install -e .[train]

or install it manually:

pip install packaging ninja
pip install "flash-attn==2.5.5" --no-build-isolation

Training WoG from Scratch

You can start the trainging from the weights of OpenVLA for greater efficiency. Please follow the instruction of OpenVLA to download their weights:

mkdir -p pretrained
cd pretrained

git clone [email protected]:openvla/openvla-7b-prismatic

cd openvla-7b-prismatic
git lfs pull

Also, you can download the pretrained ckpts(.pth) of Vision Foundation Models in pretrained/vision. The Wan VAE is available at Wan VAE. The DINO and SigLIP weights can be downloaded via utils/download_weights.py.

The data of Open X-Embodiment (OXE) can be download following OXE and OpenVLA. Then launch the training script. For one node with 8 A100 GPUs as an example:

## For Stage I:
torchrun --standalone --nnodes 1 --nproc-per-node 8 scripts/train.py \
  --pretrained_checkpoint pretrained/openvla-7b-prismatic/checkpoints/step-295000-epoch-40-loss=0.2200.pt \
  --vla.type prism-dinosiglip-224px+oxe+diffusion \
  --vla.data_mix oxe_magic_soup_plus \
  --vla.expected_world_size 8 \
  --vla.global_batch_size 256 \
  --vla.per_device_batch_size 32 \
  --vla.learning_rate 2e-5 \
  --data_root_dir <path_to_dataset_dir> \
  --run_root_dir <path_to_log/checkpoint_dir> \                 
  --run_id <optional_run_id_for_wandb> \
  --image_aug <True_or_False> \
  --wandb_project <your_wandb_project> \
  --wandb_entity <your_wandb_entity> \
  --save_interval <num_of_steps_to_save_checkpoint> \
  --repeated_diffusion_steps 8 \
  --Ta 16 \
  --action_model_type DiT-L \
  --is_resume False \
  --train_venc True 

## For Stage II:
torchrun --standalone --nnodes 1 --nproc-per-node 8 scripts/train.py \
  --pretrained_checkpoint <ckpt predtrained in stage I> \
  --vla.type prism-dinosiglip-224px+oxe+diffusion \
  --vla.data_mix oxe_magic_soup_plus \
  --vla.expected_world_size 8 \
  --vla.global_batch_size 256 \
  --vla.per_device_batch_size 32 \
  --vla.learning_rate 2e-5 \
  --data_root_dir <path_to_dataset_dir> \
  --run_root_dir <path_to_log/checkpoint_dir> \                 
  --run_id <optional_run_id_for_wandb> \
  --image_aug <True_or_False> \
  --wandb_project <your_wandb_project> \
  --wandb_entity <your_wandb_entity> \
  --save_interval <num_of_steps_to_save_checkpoint> \
  --repeated_diffusion_steps 8 \
  --Ta 16 \
  --action_model_type DiT-L \
  --is_resume False

Pretrained Checkpoints

We provide pretrained checkpoints of the two stages in huggingface. WoG-V is the first-stage and WoG-A is the second.

Evaluation in SIMPLER

In this section, we provide a minimal evaluation for our models in SIMPLER. First, please follow the instruction of SimplerEnv to install the simulation environment. Next, add our ./deploy to SimplerEnv/simpler_env/policies.

cp ./deploy <your_path_to_simpler>/simpler_env/policies -r

Then add a new policy model in SimplerEnv/simpler_env/main_inference.py as below:

elif args.policy_model == "wog":
    from simpler_env.policies.deploy import WoGSimInference
    assert args.ckpt_path is not None
    model = WoGSimInference(
        saved_model_path=args.ckpt_path,  
        policy_setup=args.policy_setup,
        action_scale=args.action_scale,
        action_model_type='DiT-L',            
    )

After that, you can modify and launch the scripts in deploy/scripts like:

cd <your_path_to_simpler>
bash simpler_env/policies/deploy/scripts/wog_put_in_drawer_visual_matching.sh

Ps: For Calvin Deployment

We now support CALVIN deployment. You can convert the CALVIN dataset to RLDS format with calvin_rlds_builder. We've registered the calvin dataset in:

• prismatic/vla/datasets/rlds/oxe/transforms.py
• prismatic/vla/datasets/rlds/oxe/mixtures.py
• prismatic/vla/datasets/rlds/oxe/configs.py

for training. For deployment:

mv deploy/wog_policy_calvin.py your_path_to_calvin/calvin_models/calvin_agent/evaluation/ 

Evaluation in RealWorld

We provide a sample of RealWorld Deploy wrapper in deploy/wog_policy_real.py. It's supported by Maniunicon and you can also use it in other platforms.

ManiUnicon is a universal real-world robot control platform for data-collection and model deploy. You can refer the README to collect data in rlds format.

Once you have collected rlds dataset, modify the following files in our project:

• prismatic/vla/datasets/rlds/oxe/transforms.py
• prismatic/vla/datasets/rlds/oxe/mixtures.py
• prismatic/vla/datasets/rlds/oxe/configs.py

Then only stage-II is needed for training:

## For RealWorld:
torchrun --standalone --nnodes 1 --nproc-per-node 8 scripts/train.py \
  --pretrained_checkpoint <ckpt predtrained in stage I> \
  --vla.type prism-dinosiglip-224px+oxe+diffusion \
  --vla.data_mix real_dataset_mix \
  --vla.expected_world_size 8 \
  --vla.global_batch_size 256 \
  --vla.per_device_batch_size 32 \
  --vla.learning_rate 2e-5 \
  --data_root_dir <path_to_dataset_dir> \
  --run_root_dir <path_to_log/checkpoint_dir> \                 
  --run_id <optional_run_id_for_wandb> \
  --image_aug <True_or_False> \
  --wandb_project <your_wandb_project> \
  --wandb_entity <your_wandb_entity> \
  --save_interval <num_of_steps_to_save_checkpoint> \
  --repeated_diffusion_steps 8 \
  --Ta 16 \
  --action_model_type DiT-L \
  --is_resume False

After training, You can follow wog config in maniunicon and wog class in maniunicon to deploy WoG.

🔗 Citation

@article{WoG,
        title={World Guidance: World Modeling in Condition Space for Action Generation}, 
        author={Yue Su and Sijin Chen and Haixin Shi and Mingyu Liu and Zhengshen Zhang and Ningyuan Huang and Weiheng Zhong and Zhengbang Zhu and Yuxiao Liu and Xihui Liu},
        journal={arXiv preprint arXiv:2602.22010},
        year={2026},
  }

License

All the code, model weights, and data are licensed under MIT license.