DrivingGPT: Unifying Driving World Modeling and Planning with Multi-modal Autoregressive Transformers

CVPR 25 2025 65.6 method, application

TLDR

Unifies driving world modeling and planning via autoregressive next-token prediction on interleaved image and action tokens.

Reasoning

The paper introduces a novel multimodal autoregressive approach that jointly handles world modeling and planning, outperforming baselines on nuPlan and NAVSIM. Strengths include unifying simulation and planning in a single framework; weaknesses include limited discussion of generalization beyond driving and potential scalability issues.

Read-first score

Read-first score 65.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

81.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 299.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Unifying driving world modeling and planning into a single sequence modeling problem
Introducing a multimodal driving language based on interleaved image and action tokens
Using autoregressive transformers with standard next-token prediction for joint world modeling and planning

Methodology

DrivingGPT employs a multimodal autoregressive transformer that processes interleaved image and action tokens as a unified sequence. The model is trained via standard next-token prediction on large-scale driving data, enabling both action-conditioned video generation and end-to-end trajectory planning.

Key Results

DrivingGPT outperforms strong baselines on the nuPlan and NAVSIM benchmarks in both action-conditioned video generation and end-to-end planning tasks.

Resources

Paper Project