Metraipolis › archive › dc_faf72d9985da
Reactive Real-Time Flow Policies via Asynchronous Distribution Alignment — arXiv:2609.36540
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.36540
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-07T17:32:39.987Z
- SHA-256 of the archived text
158247feb722a0c0eb803e0d24bd4da703ee481c0f4f0f91cafceede3ddaff8a- Stable URL of this page
https://metraipolis.com/documento/dc_faf72d9985da
How to cite
Reactive Real-Time Flow Policies via Asynchronous Distribution Alignment — arXiv:2609.36540 — arXiv (en), https://arxiv.org/abs/2609.36540 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_faf72d9985da (Metraipolis archive dc_faf72d9985da, sha256 158247feb722a0c0eb803e0d24bd4da703ee481c0f4f0f91cafceede3ddaff8a)
BibTeX
@misc{dcfaf72d9985da,
title = {Reactive Real-Time Flow Policies via Asynchronous Distribution Alignment — arXiv:2609.36540},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.36540},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_faf72d9985da; sha256 158247feb722a0c0eb803e0d24bd4da703ee481c0f4f0f91cafceede3ddaff8a},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.36540. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Generalist robot policies such as vision-language-action models (VLAs) have achieved remarkable generalization, but their inference delays can conflict with the demands of real-time control. Asynchronous execution avoids pauses between action chunks by predicting the next sequence of actions while the robot carries out the previous one. In this paper, we study whether asynchronous execution produces the same action distribution as the original VLA. We find that, for non-Markovian demonstrations, asynchronous execution can produce a fundamentally different action distribution, which can limit the policy's reactivity. In our method, we seek to restore this reactivity by aligning the asynchronously produced action distribution with that of the original VLA through two complementary mechanisms. First, Recursive Flow-Field Distillation trains the asynchronous policy using the VLA's action-generation flow. We characterize the learned distribution theoretically and show experimentally that our asynchronous policy can generate nearly the full range of actions the original VLA would produce, while existing asynchronous methods recover only a fraction of that range. Second, Propose-Resolve prepares multiple action sequences asynchronously and uses the latest observation to select among them based on a lightweight approximation of their likelihood under the VLA's action distribution. Our resulting method matches the original VLA's success on LIBERO and retains about 80% of its success on RoboMimic, about 30 percentage points more than existing asynchronous methods. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-30 in cs.LG). Categorie arXiv: Robotics (cs.RO) ; Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2609.36540 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-07T16:59:48.501Z dalla pagina https://arxiv.org/catchup/cs.LG/2026-09-30?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).