Gaussian Splatting Visual MPC for Granular Media Manipulation


ICRA 2025 Submission
Wei-Cheng Tseng, Ellina Zhang, Krishna Murthy Jatavallabhula, Florian Shkurti


Overview


Image

Our method takes a few multi-view images of a scene and their corresponding camera poses as input (a) converts them into their Gaussian splatting representation, (b) learns a dynamics model over these representations, and (c) performs visual model-predictive control for granular material manipulation, which requires view synthesis and dynamics rollouts.




Method


Image

Our method contains the following components. (a) The dynamics model f, conditioned on the input action, predicts the temporal evolution of the scene representation. During planning time, we calculate the task objective and backpropagate the gradients to optimize the action sequence (b) The dynamics model f, conditioned on the input action , predicts the temporal evolution of the scene representation . During planning time, we calculate the task objective and backpropagate the gradients to optimize the action sequence such that the granular media can get closed to the target observations.




Novel-view Synthesis Results


Our framework represent the granular material and the scene with Gaussian Splatting, which enables us do novel-view synthesis.

Image



Real-World Manipulation Results


Experimental Setup: The robotic manipulator, equipped with a pusher at the end-effector, moves object piles within the workspace. Four calibrated RGBD cameras mounted around the workspace provide visual observations of the environment. We show the results of following tasks. Splitting: pushing the piles into multiple target regions. Collecting: pushing the piles into a target region.

Qualitative Results: We show the qualitative results below. From the results we got, granular material finally match the pattern of the target observation after manipulated from their initial configuration.

Image


Manipulation Visualization: We show the manipulation recording here. After several pushes of manipulator, the granular media can reach the target region.

Image
Image




Resource and Citation


Video  Paper   Summplmentary   Code (comming soon)   Dataset (comming soon)  
  @inproceedings{icra_submission_tseng,
      author = {Wei-Cheng Tseng, Ellina Zhang, Krishna Murthy Jatavallabhula, Florian Shkurti},
      title = {Granular Material Manipulation with Gaussian Splatting},
      year = {2024}
  }