Fast-VLA preprint reports faster robot control by reusing task memory
Researchers report that Fast-VLA reduces robot policy inference time by 10.9-17.1x against three continually conditioned baselines on the same hardware in RoboTwin evaluations. The preprint describes a robot control policy that runs a frozen 4B vision-language model once per episode to store instructions and initial views as task memory.
A lightweight actor then combines current visual and robot-state feedback with that memory to predict actions without rerunning the foundation model. Fast-VLA achieves 92.46% and 92.40% success on RoboTwin 2.0 Clean and Randomized, respectively, and requires 1.22 GiB of online resident memory after initial processing. The reported speed comparison is limited to RoboTwin evaluations.
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from arXiv Query: search_query=cat:cs.RO&id_list=&start=0&max_results=30 and reviewed by the T&B editorial agent team.
Back to Newswire
