Search papers, labs, and topics across Lattice.
This paper introduces Deform360, a comprehensive visuotactile dataset designed to enhance the modeling of deformable objects in robotic manipulation, comprising 198 objects and 1,980 interaction sequences captured through advanced multi-camera and tactile sensing technologies. The study systematically evaluates the performance of existing world models, contrasting 2D video-based approaches with 3D particle models, revealing critical insights into their respective strengths and weaknesses. Preliminary robot planning tasks demonstrate the dataset's potential for real-world applications, establishing a benchmark for future research in deformable object dynamics.
A groundbreaking dataset reveals that 3D particle models significantly outperform 2D video models in capturing the complexities of deformable object dynamics.
Predicting object dynamics (i.e., world modeling) is a fundamental challenge for robotic manipulation, and modeling deformable objects presents a particularly difficult case due to their high-dimensional state spaces and complex material properties. While current world models approach this through two distinct paradigms: learning the dynamics over the 2D pixel space or more explicit 3D geometric space. A systematic understanding of their relative strengths and limitations remains elusive due to the lack of diverse, large-scale real-world data. To address this, we present Deform360, a large-scale visuotactile dataset featuring 198 daily-life objects, 1,980 interaction sequences, and over 215 hours of observations from 41 surround-view cameras and bimanual tactile grippers to capture both global motion and contact-induced local deformations. Leveraging a novel markerless visuotactile 3D tracking pipeline to extract dense geometry and motion, we systematically evaluate current state-of-the-art world models, comparing 2D video models against 3D particle models. Finally, we provide a preliminary demonstration indicating the real-world applicability of our dataset by performing robot planning tasks on deformable objects. Our analysis reveals key insights into the trade-offs between structural priors and scalability, providing a solid benchmark for future research in generalizable deformable object-centric world modeling. Project website: https://deform360.lhy.xyz
MIT CSAIL