CausalDrive: Real-time Causal World Models for Autonomous Driving

arXiv 2026 65.4 method, application

TLDR

CausalDrive is a real-time causal world model for autonomous driving that generates interactive video from ego trajectory and text prompts.

Reasoning

The paper presents a novel architecture achieving real-time 12 FPS video generation with causal prediction, enabling controllable simulation. However, the abstract is cut off and lacks explicit real-world validation or detailed evaluation metrics.

Read-first score

Read-first score 65.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Recency 6%

100

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

Citation impact 18%

93.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93913108

Topical relevance 29%

87.1

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=evaluation,experiment

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 477.

Keyword Scores

world model

interactive world model

world simulator

generative world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Operates solely on initial front-view frame, ego trajectory, and text prompt, excluding future NPC layouts to force causal interaction prediction
Text-driven control over Driving Sociology, enabling dynamic orchestration of counterfactual reactions to identical ego-actions
Context-Forced DMD architecture combining continuous flow-matching with a self-correcting distillation objective for real-time generation at 12 FPS
Transforms passive video generator into a playable neural simulator
Three downstream applications: generative closed-loop evaluation, RL post-training with Video2Reward, and real-time human-in-the-loop simulation

Methodology

CausalDrive uses only the initial front-view frame, ego-vehicle's trajectory, and a macroscopic text prompt as input, without future NPC layouts, to force causal interaction prediction. It employs a Context-Forced DMD architecture that combines continuous flow-matching with a self-correcting distillation objective to achieve real-time generation at 12 FPS. The model is evaluated on three downstream tasks: closed-loop evaluation, RL post-training via Video2Reward, and human-in-the-loop simulation.

Key Results

CausalDrive achieves interactive speeds of 12 FPS, significantly mitigating collision artifacts in closed-loop evaluation. Policies trained within its reactive scenarios exhibit superior interaction capabilities in the real world.

Resources

Paper