UniWorld: Autonomous Driving Pre-training via World Models

arXiv 23.8 2023 62 method, application

TLDR

UniWorld uses a spatial-temporal world model for label-free pre-training on image-LiDAR data, improving autonomous driving tasks like motion prediction and 3D detection.

Reasoning

The paper presents a novel pre-training framework based on 4D occupancy prediction, showing clear improvements on nuScenes benchmarks and reducing annotation costs. However, it lacks comparisons to other world model methods and real-world deployment validation.

Read-first score

Read-first score 62, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 30.

Reproducibility 25%

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

Methodology quality 25%

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

Recency 8%

65.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2023

Topical relevance 42%

42.9

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

Field roles

Methodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 489.

Keyword Scores

world model

world dynamics prediction

generative world model

world simulator

video world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Drawing inspiration from occupancy grid as World Models for robots, introducing spatial-temporal world model UniWorld
Label-free pre-training process enabling use of massive image-LiDAR pairs
Unified pre-training framework that predicts 4D geometric occupancy as World Models and fine-tunes on multiple downstream tasks
Ability to estimate missing world state information and predict plausible future states

Methodology

UniWorld first predicts 4D geometric occupancy as World Models in a foundational stage, then fine-tunes on downstream tasks. The pre-training is label-free and uses massive amounts of image-LiDAR pairs. The framework is evaluated on motion prediction, multi-camera 3D object detection, and surrounding semantic scene completion.

Key Results

On nuScenes dataset, UniWorld improves motion prediction IoU by 1.5%, multi-camera 3D detection mAP and NDS by 2.0% each, and semantic scene completion mIoU by 3%. It also achieves a 25% reduction in 3D training annotation costs.

Resources

Paper Code