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

WorldRFT: Latent World Model Planning with Reinforcement Fine-Tuning for Autonomous Driving

AAAI 26 2026 49.2 method

TLDR

WorldRFT proposes a planning-oriented latent world model with reinforcement fine-tuning for autonomous driving, achieving SOTA with 83% collision reduction.

Reasoning

The paper introduces a novel framework combining hierarchical planning decomposition and reinforcement fine-tuning, demonstrating strong empirical results on nuScenes and NavSim. However, the approach is domain-specific and its generalizability to other tasks is not explored.

Read-first score

Read-first score 49.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 8%
100

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

Topical relevance 42%
50

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%
50

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

Reproducibility 25%
30

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 518.

Keyword Scores

world model
10
model-based reinforcement learning world model
8
world dynamics prediction
6
interactive world model
5
world simulator
3
generative world model
2
video world model
1

Deep Analysis

Innovations

  • Planning-oriented latent world model framework that aligns scene representation learning with planning via hierarchical planning decomposition and local-aware interactive refinement mechanism
  • Reinforcement fine-tuning (RFT) with Group Relative Policy Optimization (GRPO) using trajectory Gaussianization and collision-aware rewards
  • Integration of a vision-geometry foundation model to improve 3D spatial awareness

Methodology

WorldRFT integrates a vision-geometry foundation model for 3D spatial awareness, employs hierarchical planning task decomposition to guide representation optimization, and uses local-aware iterative refinement to derive a planning-oriented driving policy. It then applies Group Relative Policy Optimization (GRPO) with trajectory Gaussianization and collision-aware rewards to fine-tune the policy for safety-critical performance.

Key Results

WorldRFT achieves state-of-the-art performance on nuScenes, reducing collision rates by 83% (from 0.30% to 0.05%), and on NavSim with camera-only sensors it attains competitive performance (87.8 PDMS) compared to the LiDAR-based SOTA method DiffusionDrive (88.1 PDMS).

Resources

Paper

Tags