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

Occupancy World Model for Robots

arXiv 25.5 2025 57.7 method

TLDR

Proposes RoboOccWorld, an occupancy world model for indoor scene evolution forecasting using spatio-temporal attention and conditional causal state attention.

Reasoning

Strengths include addressing indoor scene forecasting with novel components (CCSA, HSTA) and a new benchmark. Weaknesses are the narrow focus on occupancy prediction and lack of real robot deployment evidence.

Read-first score

Read-first score 57.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Methodology quality 25%
90

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

Recency 8%
86.7

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

Topical relevance 42%
44.3

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 25%
38

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 380.

Keyword Scores

world model
8
world dynamics prediction
8
generative world model
7
world simulator
4
video world model
2
interactive world model
1
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • Proposes RoboOccWorld, an occupancy world model for indoor 3D scene evolution forecasting using a combined spatio-temporal receptive field and guided autoregressive transformer.
  • Introduces Conditional Causal State Attention (CCSA) that uses camera poses of the next state as conditions to guide the autoregressive transformer for indoor robotics scenarios.
  • Proposes Hybrid Spatio-Temporal Aggregation (HSTA) to obtain a combined spatio-temporal receptive field based on multi-scale spatio-temporal windows.
  • Restructures the OccWorld-ScanNet benchmark with local annotations to facilitate evaluation of indoor 3D occupancy scene evolution prediction.

Methodology

RoboOccWorld employs a guided autoregressive transformer with Conditional Causal State Attention (CCSA) that conditions on camera poses of the next state to adapt to indoor robotics scenarios. It also uses Hybrid Spatio-Temporal Aggregation (HSTA) to extract multi-scale spatio-temporal cues from historical observations via combined spatio-temporal receptive fields. The model is evaluated on a restructured OccWorld-ScanNet benchmark for indoor 3D occupancy scene evolution prediction.

Key Results

RoboOccWorld outperforms state-of-the-art methods on the indoor 3D occupancy scene evolution prediction task, as demonstrated on the restructured OccWorld-ScanNet benchmark.

Limitations

  • The method is evaluated only on indoor scenes (ScanNet) and may not generalize to outdoor or unstructured environments.
  • The approach relies on camera poses as conditions for the autoregressive transformer, which may not always be available or accurate in real-world robotics settings.
  • The benchmark restructuring uses local annotations, which could limit applicability to scenes without such detailed annotations.

Resources

Paper

Tags