Elamur Research⌕
☰
home / papers / 2026 / Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System

Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System

Paper · 2026 · Yifei Wei, Linqing Zhong, Yi Liu et al.

Summary

Libra-VLA addresses a weakness of most Vision-Language-Action (VLA) robot policies. These policies translate visual and linguistic features straight into continuous motor commands in one flat step, which leaves a wide gap between semantics and actuation. The authors instead treat actions as living in a Hybrid Action Space. A coarse discrete component encodes the general direction of motion, and a continuous component adds fine pose corrections.

The system has two parts that run at different speeds:

  • Semantic Planner (System 2): a VLM with a parallel decoding head that predicts coarse action tokens at low frequency.
  • Action Refiner (System 1): a diffusion transformer with its own high-resolution visual encoder that produces continuous residuals at high frequency, conditioned on the planner's tokens.

An intent buffer lets the two modules run asynchronously. Training uses a dynamic teacher-forcing curriculum.

The key finding is that success rate follows an inverted-U curve as coarse-bin granularity varies. It peaks when the learning difficulty is split evenly between the two modules, which the authors call the "Libra point." The design yields high benchmark success with lower latency, which matters for responsive real-world manipulation.

Key points

  • The "Libra point" is at N = 10 coarse bins. Coarser bins give too little guidance, while N ≥ 50 overloads the planner and its errors cascade into the refiner.
  • The Horizon Expansion Factor M controls how often the planner updates. Raising M from 2 to 5 lowers average success only slightly (97.2% → 95.3%) but cuts latency from 122 to 104 ms. Against the 220 ms Libra-Base baseline on an RTX 4090, this is a 44.5–52.7% reduction.
  • The dynamic curriculum beats both pure teacher forcing and no teacher forcing. It starts by conditioning on ground-truth coarse actions and later switches to predicted ones.
  • Unlike GR00T N1, FiS-VLA, OpenHelix or HybridVLA, the two modules communicate through explicit, interpretable coarse actions rather than opaque latent vectors.
  • In real-world tests, Libra-VLA was compared with π0 and Go-1 on three long-horizon tasks: Wipe Stain, Pour Water and Make Sandwich.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Beihang UniversityBeijing · China65 works · 8,452 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Yifei Wei, Linqing Zhong, Yi Liu, Yuxiang Lu, Xindong He, Maoqing Yao and Guanghui Ren. 2026. Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1848

@inproceedings{wei2026libra,
  title = {{Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System}},
  author = {Yifei Wei, Linqing Zhong, Yi Liu, Yuxiang Lu, Xindong He, Maoqing Yao and Guanghui Ren},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1848},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1848}
}

References

39 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. CoRR, abs/2502.13923…
  2. Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, and 22 others. 2025. GR00T N1: an open foundation m… find
  3. Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, and 5 others. 2024. π 0 :… find
  4. Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, Yuxin Jiang, Cheng Jing, Hongyang Li, Jialu Li, Chiming Liu, Yi Liu, Yuxiang Lu, Jianlan Luo, Ping Luo, and 31 others. 2025a. Agibot world colosseo: A large-scale manipulation platform f… find
  5. Qingwen Bu, Yanting Yang, Jisong Cai, Shenyuan Gao, Guanghui Ren, Maoqing Yao, Ping Luo, and Hongyang Li. 2025b. Univla: Learning to act anywhere with task-centric latent actions. CoRR, abs/2505.06111. — [arXiv:2505.06111](https://arxiv.org/abs/2505.06111) find
  6. Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, Deli Zhao, and Hao Chen. 2025. Worldvla: Towards autoregressive action world model. CoRR, abs/2506.21539. — [arXiv:2506.21539](https://arxiv.org/abs/2506.21539) find
  7. Hao Chen, Jiaming Liu, Chenyang Gu, Zhuoyang Liu, Renrui Zhang, Xiaoqi Li, Xiao He, Yandong Guo, Chi-Wing Fu, Shanghang Zhang, and Pheng-Ann Heng. 2025. Fast-in-slow: A dual-system foundation model unifying fast manipulation within slow reasoning. CoRR, abs/2506.01953. — [arXiv:2506.01953](https://arxiv.org/abs/2506.01953) find
  8. Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. 2025. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 44(10-11):1684– 1704. find
  9. Can Cui, Pengxiang Ding, Wenxuan Song, Shuanghao Bai, Xinyang Tong, Zirui Ge, Runze Suo, Wanqi Zhou, Yang Liu, Bofang Jia, Han Zhao, Siteng Huang, and Donglin Wang. 2025. Openhelix: A short survey, empirical analysis, and open-source dual-system vla model for robotic manipulation. Preprint, arXiv:25… — [arXiv:2505.03912](https://arxiv.org/abs/2505.03912) find
  10. Jonathan St. B. T. Evans. 2008. Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol., 59(1):255–278. find
  11. Kuan Fang, Fangchen Liu, Pieter Abbeel, and Sergey Levine. 2024. MOKA: open-world robotic manipulation through mark-based visual prompting. In Robotics: Science and Systems XX, Delft, The Netherlands, July 15-19, 2024. find
  12. Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. 2025. Libero-plus: In-depth robustness analysis of vision-language-action models. CoRR, abs/2510.13626. — [arXiv:2510.13626](https://arxiv.org/abs/2510.13626) find
  13. Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. 2023. Look before you leap: Unveiling the power of GPT-4V in robotic vision-language planning. CoRR, abs/2311.17842. — [arXiv:2311.17842](https://arxiv.org/abs/2311.17842) find
  14. Chia-Yu Hung, Qi Sun, Pengfei Hong, Amir Zadeh, Chuan Li, U-Xuan Tan, Navonil Majumder, and Soujanya Poria. 2025. NORA: A small open-sourced generalist vision language action model for embodied tasks. CoRR, abs/2504.19854. — [arXiv:2504.19854](https://arxiv.org/abs/2504.19854) find
  15. Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, and 17 others. 2… find
  16. Yiding Jiang, Shixiang Gu, Kevin Murphy, and Chelsea Finn. 2019. Language as an abstraction for hierarchical deep reinforcement learning. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vanco… find
  17. Daniel Kahneman. 2011. Thinking, fast and slow. Farrar, Straus and Giroux. find
  18. Moo Jin Kim, Chelsea Finn, and Percy Liang. 2025. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645. find
  19. Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Paul Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. 2024. Openvla: An open-so… find
  20. Yi Li, Yuquan Deng, Jesse Zhang, Joel Jang, Marius Memmel, Raymond Yu, Caelan Reed Garrett, Fabio Ramos, Dieter Fox, Anqi Li, Abhishek Gupta, and Ankit Goyal. 2025. HAMSTER: hierarchical action models for open-world robot manipulation. CoRR, abs/2502.05485. — [arXiv:2502.05485](https://arxiv.org/abs/2502.05485) find
  21. Zhixuan Liang, Yizhuo Li, Tianshuo Yang, Chengyue Wu, Sitong Mao, Liuao Pei, Xiaokang Yang, Jiangmiao Pang, Yao Mu, and Ping Luo. 2025. Discrete diffusion VLA: bringing discrete diffusion to action decoding in vision-language-action policies. CoRR, abs/2508.20072. — [arXiv:2508.20072](https://arxiv.org/abs/2508.20072) find
  22. Yue Liao, Pengfei Zhou, Siyuan Huang, Donglin Yang, Shengcong Chen, Yuxin Jiang, Yue Hu, Jingbin Cai, Si Liu, Jianlan Luo, Liliang Chen, Shuicheng Yan, Maoqing Yao, and Guanghui Ren. 2025. Genie en-visioner: A unified world foundation platform for robotic manipulation. CoRR, abs/2508.05635. — [arXiv:2508.05635](https://arxiv.org/abs/2508.05635) find
  23. Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. 2023. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36:44776–44791. find
  24. Jiaming Liu, Hao Chen, Pengju An, Zhuoyang Liu, Renrui Zhang, Chenyang Gu, Xiaoqi Li, Ziyu Guo, Sixiang Chen, Mengzhen Liu, Chengkai Hou, Mengdi Zhao, KC alex Zhou, Pheng-Ann Heng, and Shanghang Zhang. 2025. Hybridvla: Collaborative diffusion and autoregression in a unified vision-language-action mo… — [arXiv:2503.10631](https://arxiv.org/abs/2503.10631) find
  25. Qi Lv, Weijie Kong, Hao Li, Jia Zeng, Zherui Qiu, Delin Qu, Haoming Song, Qizhi Chen, Xiang Deng, and Jiangmiao Pang. 2025. F1: A vision-language-action model bridging understanding and generation to actions. arXiv preprint arXiv:2509.06951. — [arXiv:2509.06951](https://arxiv.org/abs/2509.06951) find
  26. Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. 2025. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747.
  27. Yide Shentu, Philipp Wu, Aravind Rajeswaran, and Pieter Abbeel. 2024. From llms to actions: Latent codes as bridges in hierarchical robot control. CoRR, abs/2405.04798. — [arXiv:2405.04798](https://arxiv.org/abs/2405.04798) find
  28. Lucy Xiaoyang Shi, Brian Ichter, Michael Equi, Liyiming Ke, Karl Pertsch, Quan Vuong, James Tanner, Anna Walling, Haohuan Wang, Niccolo Fusai, Adrian Li-Bell, Danny Driess, Lachy Groom, Sergey Levine, and Chelsea Finn. 2025. Hi robot: Open-ended instruction following with hierarchical vision-languag… — [arXiv:2502.19417](https://arxiv.org/abs/2502.19417) find
  29. Martin Stolle and Doina Precup. 2002. Learning options in reinforcement learning. In Abstraction, Reformulation, and Approximation, pages 212–223, Berlin, Heidelberg. Springer Berlin Heidelberg. find
  30. LIBERO-Spatial: This suite contains 10 tasks where the robot must perform the same manipulation primitive but operates under varying spatial layouts. The objects remain consistent, but their relative positions change, requiring the agent to possess robust spatial reasoning capabilities. find
  31. Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. 2024. Octo: An open-source… find
  32. Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986.
  33. LIBERO-Object: Comprising 10 tasks, this suite fixes the spatial layout and task structure but introduces diverse object instances. The agent must generalize its manipulation skills across different visual textures and geometries, testing its object-centric visual grounding. find
  34. Chuheng Zhang, Rushuai Yang, Xiaoyu Chen, Kaixin Wang, Li Zhao, Yi Chen, and Jiang Bian. 2025a. How do vlas effectively inherit from vlms? Preprint, arXiv:2511.06619. — [arXiv:2511.06619](https://arxiv.org/abs/2511.06619) find
  35. Wenyao Zhang, Hongsi Liu, Zekun Qi, Yunnan Wang, Xinqiang Yu, Jiazhao Zhang, Runpei Dong, Jiawei He, He Wang, Zhizheng Zhang, Li Yi, Wenjun Zeng, and Xin Jin. 2025b. Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge. CoRR, abs/2507.04447. — [arXiv:2507.04447](https://arxiv.org/abs/2507.04447) find
  36. LIBERO-Goal: This suite includes 10 tasks that share the same workspace and object arrangement but differ in the semantic goals. This evaluates the agent’s ability to follow and execute distinct instructions within an identical visual context. find
  37. Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, Ankur Handa, Tsung-Yi Lin, Gordon Wetzstein, Ming-Yu Liu, and Donglai Xiang. 2025. Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. In IEEE/CVF Conf… find
  38. LIBERO-Long: As the most challenging suite, it consists of 10 long-horizon tasks. Each task requires the sequential execution of multiple primitives to complete a complex objective. This suite rigorously tests the model’s ability to handle temporal dependencies and multi-stage planning. find
  39. Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. 2023. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wei-2026-libra-vla-achieving-learning-equilibrium/