Zefu Lin, Rongxu Cui, Junjia Xu, Xiaojuan Jin, Wenling Li, Lue Fan, and Zhaoxiang Zhang
Overview • News • Documentation • Citation • Acknowledgements
World Pilot steers a VLA with priors from a World-Action Model. VLA methods generate actions from a VLM's encoding of the scene. World Pilot adds two priors from a WAM into the decision chain, with Latent Steering routing a scene-evolution latent into VLM hidden states and Action Steering feeding a trajectory-level motion prior to the action generator. This gives the VLA an anticipated view of the scene and a motion hint alongside its semantic conditioning. World Pilot reaches state-of-the-art performance on LIBERO-Plus and real-robot tasks.
- [2026.6.9] WorldPilot model weights are now available on Hugging Face. Feel free to try them out! 🚀
- [2026.6.8] The WorldPilot codebase is released, including training and evaluation. Model weights are coming soon. 🚀
- [2026.6.7] WorldPilot is now live on arXiv. The code is coming soon. 🚀
The documentation below covers environment setup, training, and public evaluation. Start with the installation guide, then follow the training or evaluation notes for your workflow.
We release our pretrained model parameters and precomputed LIBERO cache on Hugging Face.
- Model weights: Chedan86/WorldPilot-LIBERO
- Precomputed LIBERO cache: Chedan86/WorldPilot-LIBERO-precompute
If WorldPilot helps your research, we would appreciate a citation using the BibTeX entry below.
@article{worldpilot2026,
title={World Pilot: Steering Vision-Language-Action Models with World-Action Priors},
author={Zefu Lin and Rongxu Cui and Junjia Xu and Xiaojuan Jin and Wenling Li and Lue Fan and Zhaoxiang Zhang},
journal={arXiv preprint arXiv:2606.12403},
year={2026}
}We sincerely thank the teams behind ABot-Manipulation, cosmos-policy, LIBERO, LIBERO-plus, LeRobot for their outstanding work.
