SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model

arXiv 25.11 2025 34.2 method, application

TLDR

A transformer-based sparse occupancy world model for trajectory-conditioned future 3D scene occupancy forecasting, achieving SOTA on nuScenes.

Reasoning

The paper presents a novel architecture that avoids VAE tokenization and BEV projection, directly predicting multi-frame occupancy from raw images. Its strength lies in the end-to-end transformer design and strong empirical results on a real-world benchmark. Weaknesses include lack of discussion on generalization beyond nuScenes and no comparison to generative or interactive world models.

Read-first score

Read-first score 34.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.

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=benchmark,metric

Topical relevance 29%

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=code

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

Frontier

Rank sensitivity

Stability: volatile; rank range: 130.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

End-to-end prediction of multi-frame future occupancy directly from raw image features, avoiding VAE-based discrete tokenization
Sparse occupancy representation that bypasses intermediate bird's eye view (BEV) projection and explicit geometric priors
Attention-based transformer architecture inspired by GPT and VGGT to capture spatiotemporal dependencies effectively
Trajectory-conditioned forecasting enabling robust scene dynamics understanding under arbitrary future trajectories

Methodology

The method predicts multi-frame future 3D scene occupancy in an end-to-end manner directly from raw image features using a sparse occupancy representation and an attention-based transformer architecture. It avoids both the finite-capacity constraint of discrete tokenization via VAEs and the structural limitations of BEV projection.

Key Results

Achieves state-of-the-art performance on the nuScenes benchmark for 1-3 second occupancy forecasting, outperforming existing approaches by a significant margin, and demonstrates robust scene dynamics understanding under arbitrary future trajectory conditioning.

Resources

Paper