ReSim: Reliable World Simulation for Autonomous Driving

arXiv 25.6 2025 47.5 method, application

TLDR

ReSim enriches real-world driving data with simulator non-expert data to build a controllable world model for reliable simulation of diverse scenarios.

Reasoning

Strengths: addresses data scarcity of non-expert behaviors, uses heterogeneous corpus, achieves significant improvements in fidelity and controllability. Weaknesses: relies on simulator data which may have domain gap; Video2Reward module adds complexity.

Read-first score

Read-first score 47.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 59.

Recency 6%

86.7

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

Topical relevance 29%

84.3

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

Methodology quality 18%

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 578.

Keyword Scores

world model

world simulator

generative world model

video world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Enriching real-world human demonstrations with diverse non-expert data collected from a driving simulator (e.g., CARLA) to enable simulation of hazardous or non-expert behaviors
Building a controllable world model using a diffusion transformer architecture with strategies to integrate conditioning signals for improved prediction controllability and fidelity
Introducing a Video2Reward module that estimates a reward from simulated future frames to bridge the gap between high-fidelity simulation and applications requiring reward signals

Methodology

ReSim uses a video generator based on a diffusion transformer architecture trained on a heterogeneous corpus combining real-world driving data and diverse non-expert data from the CARLA simulator. Several strategies are devised to effectively integrate conditioning signals, enhancing controllability and fidelity. The model is evaluated on visual fidelity, controllability for both expert and non-expert actions, and planning/policy selection performance on the NAVSIM benchmark.

Key Results

ReSim achieves up to 44% higher visual fidelity, improves controllability for both expert and non-expert actions by over 50%, and boosts planning and policy selection performance on NAVSIM by 2% and 25%, respectively.

Resources

Paper Project