Flex-π: A Multi-Stream World-Action Model with Compute Flexibility

Paper (arXiv:2608.10860)  ·  Project page

Ge Yan*¹, Jinghao Liu*¹, Yuzhi Fan*¹, Lei Cai¹, Minwen Liao¹, Jesse Zhang†¹, Dieter Fox†^1,2

¹University of Washington  ·  ²Allen Institute for AI
*Equal contribution. †Equal advising.

The code is ready soon. We are finishing the cleanup and will push the training and inference code, along with pre-trained checkpoints, to this repository. Watch the repo to be notified when it lands.

Until then, the project page has the full set of real-robot rollouts and the paper has the architecture and evaluation details.

World-action models predict the future to act better, but they mainly predict RGB latents trained for pixel reconstruction — with no explicit signal for the 3D geometry or object semantics that manipulation needs.

Flex-π is a 6B-parameter world-action model that jointly denoises RGB, 3D geometry, and object-centric DINO semantics together with actions in a shared latent space. Per-stream dropout with cross-modality forcing yields a single checkpoint that runs on any subset of streams — from fast action-only inference to full joint generation — with no new sensors or visual priors.

The result is a demonstration-efficient policy that beats the strongest baselines by up to 2–7× on dexterous, precise, real-world bimanual tasks both in and out of distribution, while running faster than π_0.5.

@article{yan2026flexpi,
  title   = {Flex-$\pi$: A Multi-Stream World-Action Model with Compute Flexibility},
  author  = {Yan, Ge and Liu, Jinghao and Fan, Yuzhi and Cai, Lei and
             Liao, Minwen and Zhang, Jesse and Fox, Dieter},
  journal = {arXiv preprint arXiv:2608.10860},
  year    = {2026}
}