Action-expert checkpoint for GE-Act post-trained with Robust-WAM, a lightweight semantic-foresight alignment that makes video-generation World-Action Models robust to visual out-of-distribution shifts. During training, learnable query tokens prepended to the action stream are cosine-aligned to the frozen DINOv3 CLS embeddings of future frames; at inference the teacher and alignment head are dropped and only the query tokens remain, so the overhead is negligible.

📄 Paper: https://arxiv.org/abs/2608.05903
🌐 Project page: https://haodong-yan.github.io/robust-wam-project-page/
💻 Code: https://github.com/Haodong-Yan/Robust-WAM-release

What this is

Backbone: GE-Act (LTX-Video action-expert WAM).
Training: LIBERO (four standard suites), Robust-WAM alignment (bidirectional prefix, K=16 query tokens, DINOv3 ViT-B/16 CLS targets), step 50000.
Files: config.json + diffusion_pytorch_model.safetensors (the action-expert weights).

Results

On LIBERO-Plus (out-of-distribution perturbations), Robust-WAM lifts GE-Act from 78.0 → 80.9 overall success rate while keeping clean-LIBERO performance (96.5 → 97.3).

Usage

This is the action-expert checkpoint only; use it with the release code plus the base LTX-Video weights and the LIBERO-Plus benchmark.

bash

git clone https://github.com/Haodong-Yan/Robust-WAM-release
cd Robust-WAM-release/geact-RW

# point the eval config at this checkpoint's directory (holding config.json + safetensors)
export GEACT_CKPT_PATH=/path/to/this/download
export LTX_MODEL_PATH=/path/to/ltx-video
export DATA_ROOT=/path/to/data
export DINOV3_MODEL_PATH=/path/to/dinov3_vitb16   # referenced by config; teacher not run at inference
export PYTHONPATH=/path/to/LIBERO-plus:$PYTHONPATH
export MUJOCO_GL=egl PYOPENGL_PLATFORM=egl

python experiments/eval_libero_plus.py \
    --config_file eval_out/lplus/eval_lplus_bidir.yaml \
    --ckpt_path $GEACT_CKPT_PATH/diffusion_pytorch_model.safetensors \
    --out_dir outputs/eval_bidir --device 0 --shard 0 --num_shards 1

See geact-RW/README.md for full setup.

Acknowledgement

Built on GE-Act / Genie and LTX-Video; semantic teacher is DINOv3; evaluated on LIBERO and LIBERO-Plus. We thank the authors of these projects.

Citation

bibtex

@misc{yan2026robustwam,
  title         = {Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models},
  author        = {Haodong Yan and Junfeng Li and Junjie He and Zhide Zhong and MingMing Yu and Wenxuan Song and Jiaguan Zhu and Yangyang Zheng and Yuqiao Du and Jiadi You and Yingjie Cai and Xu Yan and Guanyi Zhao and Bingbing Liu and Haoang Li},
  year          = {2026},
  eprint        = {2608.05903},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.05903}
}

Haodong082399/Robust-WAM-GE-Act

Robust-WAM · GE-Act (LIBERO) checkpoint

What this is

Results

Usage

Acknowledgement

Citation

Haodong082399/
Robust-WAM-GE-Act