DIO: Decomposable Implicit 4D Occupancy-Flow World Model

CVPR 25 2025 53.3 method

TLDR

DIO is a decomposable 4D occupancy-flow world model that forecasts instance shapes from LiDAR, achieving SOTA on Argoverse 2.

Reasoning

The paper presents a novel world model with flexible instance decomposition and forecasting, achieving strong empirical results on real-world benchmarks. However, it is limited to LiDAR data and does not address interactive or generative capabilities, nor model-based RL.

Read-first score

Read-first score 53.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 27.

Citation impact 18%

98.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.9838069

Recency 6%

86.7

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

Methodology quality 18%

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

Topical relevance 29%

38.6

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

Reproducibility 18%

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

Citation velocity 12%

18.7

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=1.50

Field roles

FoundationFrontier

Rank sensitivity

Stability: volatile; rank range: 402.

Keyword Scores

world model

world dynamics prediction

world simulator

generative world model

video world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Decomposable implicit 4D occupancy-flow world model that estimates scene occupancy-flow from sparse LiDAR and decomposes it into individual instances.
Flexible prompt representation that allows integration with off-the-shelf 3D detectors for instance prompts.
Ability to both complete instance shapes at the present time and forecast their occupancy-flow evolution over a future horizon.

Methodology

DIO is a world model that takes sparse LiDAR observations, estimates scene occupancy-flow, and decomposes it into individual instances using an implicit representation. It employs a flexible prompt representation to incorporate instance prompts from off-the-shelf 3D detectors. The model is trained and evaluated on the Argoverse 2 dataset for 4D semantic occupancy completion and forecasting, and is also transferred to the task of LiDAR point cloud forecasting.

Key Results

DIO achieves state-of-the-art performance in 4D semantic occupancy completion and forecasting on the Argoverse 2 dataset, and ranks first in the Argoverse 4D occupancy forecasting challenge for LiDAR point cloud forecasting.

Resources

Paper