Awesome World Model Hub Papers · Datasets · Projects
← Back to papers

Vega: Learning to Drive with Natural Language Instructions

arXiv 26.3 2026 53.4 method, benchmark, application

TLDR

Vision-language-action models have reshaped autonomous driving to incorporate languages into the decision-making process.

Reasoning

Fallback reasoning generated from available title and abstract metadata: Vision-language-action models have reshaped autonomous driving to incorporate languages into the decision-making process. However, most existing pipelines only utilize the language modality for scene descriptions or reasoning and lack the flexibility to follow diverse user instructions for personalized...

Read-first score

Read-first score 53.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals.

Recency 8%
100

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

Reproducibility 25%
81

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

Methodology quality 25%
60

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

Topical relevance 42%
23.5

Matches configured research keywords against title, abstract, tags, and analysis text. matched=5

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 600.

Deep Analysis

Innovations

  • Construction of a large-scale driving dataset (InstructScene) with around 100,000 scenes annotated with diverse driving instructions and corresponding trajectories
  • Proposal of a unified Vision-Language-World-Action model (Vega) that combines autoregressive paradigm for vision and language with diffusion paradigm for world modeling and action
  • Use of joint attention for modality interactions and individual projection layers for different modalities to enhance capabilities

Methodology

The authors first constructed the InstructScene dataset containing approximately 100,000 driving scenes annotated with diverse instructions and trajectories. They then proposed the Vega model, which employs an autoregressive paradigm to process visual inputs and language instructions, and a diffusion paradigm to generate future predictions (world modeling) and trajectories (action). Joint attention enables interactions between modalities, and individual projection layers are used for different modalities to improve capabilities.

Key Results

Extensive experiments demonstrate that Vega achieves superior planning performance and exhibits strong instruction-following abilities, paving the way for more intelligent and personalized driving systems.

Tags