Summary
NeuROK (Neural Object Kinematics) is a paper by Chen Geng, Guangzhao He, Yue Gao, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu, posted on arXiv (2605.30347) in computer vision. The work tackles generative 4D dynamics: producing realistic, physically plausible temporal deformations of static 3D objects under varied physical conditions. Existing approaches typically assume predefined physical models and estimate parameters via system identification, limiting them to specific object categories and small datasets. NeuROK instead learns a data-driven, object-centric parametrization of kinematic states: a latent space representing all possible states of an object, plus a decoder that maps any sampled latent code to a plausible deformed shape. The authors train a Transformer-based encoder-decoder on a curated large-scale 4D dataset. Viewed through the Lagrangian mechanics lens of classical physics, generative dynamics reduces to modeling motion within the low-dimensional latent space, greatly simplifying simulation of realistic dynamics. The paper demonstrates effectiveness and generality across multiple dynamic object types, outperforming prior work.
NeuROK: Generative 4D Neural Object Kinematics
Paper Overview
- Field: Computer Vision (CV)
- Authors: Chen Geng, Guangzhao He, Yue Gao, Yunzhi Zhang, Shangzhe Wu, Jiajun Wu
- Published: 2026-05-28
- arXiv: 2605.30347
Summary
Data-driven methods have revolutionized 3D vision, enabling Transformers to effectively reconstruct and generate static 3D objects. However, generating simulatable 4D dynamics—realistic temporal deformations of static objects under various physical conditions—remains challenging and often ad hoc, despite being crucial for building comprehensive 3D world models.
Most existing methods assume a predefined physical model and estimate its parameters via system identification, restricting them to specific object categories and small-scale datasets. The authors argue these limitations can be overcome by learning a data-driven parametrization of the kinematic states of object-centric physical systems.
Specifically, NeuROK learns:
- A latent space representing all possible states of an object
- A decoder that maps any sampled latent code to a plausible deformed shape of the object
The framework learns a Transformer-based encoder-decoder model on a carefully curated, large-scale 4D dataset. From the Lagrangian mechanics perspective of classical physics, this formulation greatly simplifies generating simulatable dynamics: one only needs to model dynamics within the low-dimensional latent space.
The paper demonstrates the effectiveness and generality of this neural simulation framework across a variety of dynamic object types, clearly outperforming prior work.
---
*Originally posted on zhichai.net, auto-collected 2026-06-01.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177980672