Embodied AI Reading Notes (@EmbodiedAIRead)
2025-09-16 | โค๏ธ 180 | ๐ 34
Towards Scalable Gaussian World Models for Robotic Manipulation
Project:ย https://gaussian-world-model.github.io/ Paper:ย https://arxiv.org/abs/2508.17600v1
This paper extends image-based world models to 3D Gaussian Splatting based world model by inferring the propagation of Gaussian primitives under the effect of robot actions.
-
Overall pipeline: real-world image inputs โ 3D reconstruction model โ current step Gaussian Splats โ 3D variational encoder โ latent gaussian representation โ diffusion transformer conditioned on observations and robot actions โ next step Gaussian splats.
-
Authors demonstrate Gaussian World Model can be easily integrated into offline imitation learning and online model-based RL for policy learning in robot manipulation tasks.
-
Authors show experiments in some sim and real scenarios that this pipeline works well in performance.
๐ Related
Auto-generated bookmark