DriveVA: Video Action Models are Zero-Shot Drivers

arXiv 2026 62.7 method, application

TLDR

DriveVA is an autonomous driving world model that jointly generates future videos and action trajectories, achieving strong zero-shot planning performance on NAVSIM.

Reasoning

The paper proposes a novel world model that integrates video generation and trajectory prediction, leveraging pretrained video models for stronger generalization. Strengths include the joint video-action decoding and zero-shot cross-domain results, while weaknesses include limited evidence of comparisons with prior methods and the abstract being cut off.

Read-first score

Read-first score 62.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 15%

99.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.99938509

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,dataset,evaluation,experiment

Topical relevance 25%

74.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Citation velocity 10%

36.8

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=5.00

Graph prestige 15%

10.6

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0030

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 194.

Keyword Scores

world model

generative world model

video world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Proposes DriveVA, an autonomous driving world model that jointly decodes future visual forecasts and action sequences in a shared latent generative process.
Leverages rich motion dynamics and physical plausibility priors from pretrained large-scale video generation models.
Uses a DiT-based decoder to jointly predict future action sequences/trajectories and videos for tighter alignment between planning and scene evolution.
Introduces a video continuation strategy to strengthen long-duration rollout consistency.

Methodology

DriveVA employs a shared latent generative process with a DiT-based decoder to jointly predict future trajectories and videos, inheriting priors from pretrained large-scale video generation models. It also uses a video continuation strategy for long-duration rollout consistency. Evaluation is conducted on NAVSIM, nuScenes, and Bench2Drive/CARLA v2 against a state-of-the-art world-model-based planner.

Key Results

DriveVA achieves 90.9 PDM score on NAVSIM and reduces average L2 error and collision rate by 78.9% and 83.3% on nuScenes and 52.5% and 52.4% on Bench2Drive/CARLA v2 compared with the state-of-the-art world-model-based planner.

Resources

Paper