GWM: Towards Scalable Gaussian World Models for Robotic Manipulation

ICCV 25 2025 70 method

TLDR

GWM uses Gaussian primitives and a latent Diffusion Transformer to predict future 3D scenes for robotic manipulation, enabling improved imitation learning and model-based RL.

Reasoning

The paper introduces a novel 3D world model leveraging Gaussian splatting and a latent DiT, showing strong empirical results in both simulated and real-world robotic manipulation tasks. However, the abstract lacks detailed quantitative comparisons and discussion of limitations, making it hard to fully assess scalability and generalizability.

Read-first score

Read-first score 70, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Topical relevance 42%

87.1

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%

86.7

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 213.

Keyword Scores

world model

world dynamics prediction

world simulator

generative world model

model-based reinforcement learning world model

interactive world model

video world model

Deep Analysis

Innovations

Proposes Gaussian World Model (GWM) for robotic manipulation, reconstructing future states by inferring propagation of Gaussian primitives under robot actions.
Combines a latent Diffusion Transformer (DiT) with a 3D variational autoencoder for fine-grained scene-level future state reconstruction using Gaussian Splatting.
GWM serves dual purposes: enhancing visual representation for imitation learning via self-supervised future prediction, and acting as a neural simulator for model-based reinforcement learning.

Methodology

GWM reconstructs future states by modeling the propagation of Gaussian primitives conditioned on robot actions. It employs a latent Diffusion Transformer (DiT) integrated with a 3D variational autoencoder to enable fine-grained scene-level future state reconstruction with Gaussian Splatting. The model is trained via self-supervised future prediction and can be used for imitation learning or as a simulator for model-based reinforcement learning.

Key Results

GWM precisely predicts future scenes conditioned on diverse robot actions in both simulated and real-world experiments. Policies trained using GWM outperform state-of-the-art methods by impressive margins, demonstrating the initial data scaling potential of 3D world models.

Resources

Paper Project