VITAL turns visual motion into precise driving actions
An arXiv technical report submitted on 14 October 2026, "VITAL: From Visual Motion to Precise Action", introduces VITAL (Vision-Induced Tokens for Action Learning), an action tokenizer that pairs a visually derived motion scaffold with a compact residual encoding the trajectory error left by that scaffold. Using flow matching, the action expert jointly predicts both components from the current image, ego state and navigation command, without generating future images at inference, and the same residual coordinates support bounded reinforcement post-training that keeps the visual scaffold fixed. The authors report that with only one action denoising step VITAL reaches 90.8 PDMS on NAVSIM v1 navtest, within 0.13 points of its ten-step score at 3.29x end-to-end inference speed, and without retraining scores 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard.
The authors report that single-step action denoising comes within 0.13 points of their ten-step PDMS score while running 3.29x faster end to end; all figures are self-reported in the paper with no independent measurement supplied.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.20604 (cs)
[Submitted on 14 Oct 2026]
Title:VITAL: From Visual Motion to Precise Action
Authors:Ao Liang, Youquan Liu, Mingxiao Li, Wei Chow, Lingdong Kong, Yue Liao, Dongyue Lu, Yu Yang, Wei Yin, Runnan Chen
View a PDF of the paper titled VITAL: From Visual Motion to Precise Action, by Ao Liang and 9 other authors