VITAL
An action prediction framework for autonomous driving that derives driving action targets directly from visual motion through an action tokenizer. The action representation is decomposed into a visual-motion scaffold and a normalized residual encoding, trained with flow matching, and supports single-step inference as well as bounded reinforcement post-training.
Timeline (last 90 days)
VITAL's performance when transferred to the new NAVSIM v2 benchmark without retraining shows that the representational approach of the visual motion scaffold has a certain generalization ability, rather than being overfit to a specific dataset.
VITAL (Vision-Induced Tokens for Action Learning) is an autonomous driving action learning framework that makes visual motion the primary target of action prediction and uses a compact residual branch to handle precision correction.
VITAL's action representation is split into two parts: a visual motion scaffold and a residual encoding.
VITAL's action expert module is trained using flow matching, and at inference time it relies only on the current image, ego-vehicle state and navigation instruction, without needing to generate future images.
On the navtest benchmark of NAVSIM v1, VITAL reached 90.8 PDMS with just one action denoising step, only 0.13 points behind the score of ten-step inference, with a 3.29x improvement in end-to-end inference speed.
Without retraining, VITAL reached 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard.
VITAL's residual coordinate system supports bounded reinforcement post-training, making fine adjustments only at the execution level while keeping the visual motion scaffold fixed.
VITAL's core contribution lies in reorganizing the representational structure of action learning: first making visual motion the prediction target, then using residuals to make up the precision gap, rather than first defining a trajectory space and then fitting it.
2014年香港风投公司Deep Knowledge Ventures任命名为VITAL的AI算法为董事会成员，并赋予其对投资决策的投票权
All Facts (9)
VITAL's performance when transferred to the new NAVSIM v2 benchmark without retraining shows that the representational approach of the visual motion scaffold has a certain generalization ability, rather than being overfit to a specific dataset.
50%UnverifiedVITAL (Vision-Induced Tokens for Action Learning) is an autonomous driving action learning framework that makes visual motion the primary target of action prediction and uses a compact residual branch to handle precision correction.
50%UnverifiedVITAL's action representation is split into two parts: a visual motion scaffold and a residual encoding.
50%UnverifiedVITAL's action expert module is trained using flow matching, and at inference time it relies only on the current image, ego-vehicle state and navigation instruction, without needing to generate future images.
50%UnverifiedOn the navtest benchmark of NAVSIM v1, VITAL reached 90.8 PDMS with just one action denoising step, only 0.13 points behind the score of ten-step inference, with a 3.29x improvement in end-to-end inference speed.
50%UnverifiedWithout retraining, VITAL reached 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard.
50%UnverifiedVITAL's residual coordinate system supports bounded reinforcement post-training, making fine adjustments only at the execution level while keeping the visual motion scaffold fixed.
50%UnverifiedVITAL's core contribution lies in reorganizing the representational structure of action learning: first making visual motion the prediction target, then using residuals to make up the precision gap, rather than first defining a trajectory space and then fitting it.
50%Unverified2014年香港风投公司Deep Knowledge Ventures任命名为VITAL的AI算法为董事会成员，并赋予其对投资决策的投票权
50%