Tsukada Lab. Computer Network Group

Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving

Yun Li, Jiachen Gong, Simon Thompson, Ehsan Javanmardi, Qunli Zhang, Zifan Zeng, Shiming Liu, Peng Wang, Zixuan Guo, Manabu Tsukada

, Jul 17, 2026

Summary in three lines

  • A slow vision-language backbone updates a feature cache while a lightweight action model consumes it every driving tick.
  • This design supports 20 Hz action updates with an inference cost of about 32 ms per tick.
  • In CARLA LangAuto-Short, route completion increased from 37.0 to 94.0 and driving score from 28.8 to 32.9.

Figures from the paper

Figure 2: Asynchronous fast–slow inference
Figure 2. Asynchronous fast–slow inferenceSource: Figure 2 from Li et al., arXiv (2026) (authors’ arXiv version 2607.15621, CC BY 4.0)A slower backbone writes context to a cache and the action expert reads it at each control tick.

How to cite

Y. Li, J. Gong, S. Thompson, E. Javanmardi, Q. Zhang, Z. Zeng, S. Liu, P. Wang, Z. Guo, and M. Tsukada, “Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving,” in , 2026.

BibTeX

@unpublished{Li2026a,
  title         = {Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving},
  author        = {Yun Li and Jiachen Gong and Simon Thompson and Ehsan Javanmardi and Qunli Zhang and Zifan Zeng and Shiming Liu and Peng Wang and Zixuan Guo and Manabu Tsukada},
  year          = {2026},
  date          = {2026-07-17},
  eprint        = {2607.15621},
  archiveprefix = {arXiv},
  primaryclass  = {cs.RO},
  url           = {https://arxiv.org/abs/2607.15621},
  note          = {Preprint}
}