GigaWorld-Policy: An Efficient Action-Centered World-Action Model

World-Action Models (WAM) initialized from pre-trained video generation backbones have demonstrated remarkable potential for robot policy learning. GigaWorld-Policy is an action-centered WAM that learns 2D pixel-action dynamics while enabling efficient action decoding, with optional video generation.

• 🚀 9x faster inference compared to Motus (leading WAM baseline)
• 📈 7% higher task success rate than Motus
• 💪 95% improvement over pi-0.5 on RoboTwin 2.0

conda create -n gigaworld-policy python==3.11
conda activate gigaworld-policy

pip install ./third_party/giga-train
pip install ./third_party/giga-models
pip install ./third_party/giga-datasets

Before training, you need to compute the normalization statistics and pre-compute the T5 text embeddings for your dataset.

This will generate a norm_stats_delta.json file which is required by the policy.

python -m scripts.compute_norm_stats \
  --data_paths "/path/to/dataset_dir" \
  --output_path "/path/to/norm_stats_delta.json" \
  --embodiment_id {embodiment-id} \
  --delta-mask {delta-mask} \
  --sample-rate 1.0 \
  --action-chunk 48 \

• --embodiment_id: Check compute_norm_stats.py for the mapping from robot type to ID.
• --delta_mask: A boolean mask indicating which action dimensions are deltas (True) vs. absolute values (False).

This will pre-compute and save the T5 text embeddings for the language instructions in your dataset.

python -m scripts.compute_t5_embedding \
  --repo_id "/path/to/dataset_dir" \
  --root "/path/to/dataset_dir" \
  --wan_path "/path/to/Wan2.2-TI2V-5B" \
  --device "cuda" \
  --text_len 512 \
  --t5_folder_name "t5_embedding"

After completing the data preprocessing steps, modify the config file world_action_model/configs/example.py to point to the generated files and your model weights:

Once the data is preprocessed and the configuration is set, you can start training:

python -m scripts.train --config world_action_model.configs.example.config

We provide an inference server and a simple open-loop evaluation client. Open-loop here means we sample observations (images/state) from an offline dataset and run inference, without executing actions in a real environment to collect the next observations.

python -m scripts.inference_server \
  --model_id "/path/to/huggingface_model_dir_or_id" \
  --transformer_path "/path/to/transformer_checkpoint_dir" \
  --stats_path "/path/to/norm_stats_delta.json" \
  --t5_embedding_pkl "/path/to/t5_embedding.pt"

Optionally, add --return_images to enable video visualization during inference (videos will be saved under --vis_dir):

python -m scripts.inference_server \
  --model_id "/path/to/huggingface_model_dir_or_id" \
  --transformer_path "/path/to/transformer_checkpoint_dir" \
  --stats_path "/path/to/norm_stats_delta.json" \
  --t5_embedding_pkl "/path/to/t5_embedding.pt" \
  --return_images \
  --vis_dir "./vis"

python -m scripts.inference_client \
  --dataset_paths "/path/to/dataset_dir" \
  --save_dir "./vis"

@article{ye2026gigaworld,
  title={GigaWorld-Policy: An Efficient Action-Centered World-Action Model},
  author={Ye, Angen and Wang, Boyuan and Ni, Chaojun and Huang, Guan and Zhao, Guosheng and Li, Hao and Li, Hengtao and Li, Jie and Lv, Jindi and Liu, Jingyu and Cao, Min and Li, Peng and Deng, Qiuping and Mei, Wenjun and Wang, Xiaofeng and Chen, Xinze and Zhou, Xinyu and Wang, Yang and Chang, Yifan and Li, Yifan and Zhou, Yukun and Ye, Yun and Liu, Zhichao and Zhu, Zheng},
  journal={arXiv preprint arXiv:2603.17240},
  year={2026}
}