WHALE: Towards Generalizable and Scalable World Models for Embodied Decision-making

arXiv 24.11 2024 67.7 method

TLDR

WHALE introduces behavior-conditioning and retracing-rollout to improve generalizability and uncertainty estimation in world models for embodied decision-making.

Reasoning

The paper presents novel techniques (behavior-conditioning and retracing-rollout) that address key challenges in world model generalization and uncertainty estimation, with a scalable spatial-temporal transformer architecture. However, the evaluation is limited to simulation tasks and offline scenarios, lacking real-world validation or benchmarks.

Read-first score

Read-first score 67.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 62.

Topical relevance 42%

88.6

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

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 347.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

model-based reinforcement learning world model

world simulator

interactive world model

Deep Analysis

Innovations

Behavior-conditioning to address policy distribution shift in world models
Retracing-rollout for efficient uncertainty estimation without model ensembles
Whale-ST: a scalable spatial-temporal transformer-based world model with enhanced generalizability
Whale-X: a 414M parameter world model trained on 970K trajectories from Open X-Embodiment for real-world manipulation

Methodology

WHALE framework combines behavior-conditioning and retracing-rollout techniques to improve world model generalizability and uncertainty estimation. It is architecture-agnostic and demonstrated with a spatial-temporal transformer (Whale-ST) and a large-scale model (Whale-X) trained on 970K trajectories from Open X-Embodiment.

Key Results

Whale-ST outperforms baselines in simulation tasks on value estimation accuracy and video generation fidelity. The uncertainty estimation technique enhances model-based policy optimization in offline scenarios. Whale-X shows promising scalability and strong generalizability in real-world manipulation with minimal demonstrations.

Resources

Paper