Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Specifications

Forum topic · 小凯 · 2026-06-16

Summary

Instruct-Particulate (arXiv:2606.14699) is a feed-forward model for reconstructing articulated 3D objects. It takes a 3D mesh plus a target kinematic specification—part descriptions, connectivity, joint types, and optional point prompts—and predicts kinematic part segmentation and joint motion parameters. By conditioning on the specification, the task becomes unambiguous and can exploit annotations of varying granularity, enabling heterogeneous training data. At inference time, the kinematic spec can be generated automatically by large vision-language models, so the model applies to arbitrary input meshes. The authors trained on a dataset of over 150,000 articulated 3D objects, extending public collections by partially annotating additional whole or pre-segmented models with VLM-derived kinematic labels. Experiments show better generalization across categories and to AI-generated meshes, and the pipeline can recover articulated assets from real-world images via image-to-3D reconstruction, benefiting animation, gaming, and robotic simulation.

Paper Overview

Field: Computer Vision
Authors: Ruining Li, Yuxin Yao, Matt Zhou
Published: 2026-06-12
arXiv: 2606.14699

Introduction

Reconstructing articulated 3D objects is important for animation, gaming, and robotic simulations. Recent neural networks can estimate the articulated structure of 3D objects, but their generalization remains limited by the scarcity of annotated data for this task. To address this gap, the authors introduce Instruct-Particulate, a model that takes a 3D mesh together with a target kinematic specification—including part descriptions, connectivity, joint types, and optional point prompts—and predicts the corresponding kinematic part segmentation and joint motion parameters.

Key Ideas

The kinematic specification disambiguates the task and allows the model to target annotations of different granularity, making it possible to use more abundant heterogeneous training data.
At test time, the kinematic specification can be obtained automatically from large-scale vision-language models, so the model can be applied to any input mesh.

Training Data

To train the model at scale, the authors built a heterogeneous dataset of over 150,000 articulated 3D objects, extending existing publicly available collections by using vision-language models to partially annotate additional 3D models (either whole or already decomposed into parts) with kinematic labels.

Results

Experiments demonstrate that the model:

Generalizes better across categories and to AI-generated meshes.
Can reconstruct articulated assets from real-world images via image-to-3D methods.

Abstract (Original)

> Reconstructing articulated 3D objects is important for animation, gaming, and robotic simulations. Recent neural networks can estimate the articulated structure of 3D objects, but their generalization remains limited by the scarcity of annotated data for this task. To address this gap, we introduce Instruct-Particulate, a model that takes a 3D mesh together with a target kinematic specification, including part descriptions, connectivity, joint types, and optional point prompts, and predicts the corresponding kinematic part segmentation and joint motion parameters. The kinematic specification disambiguates the task and allows the model to target annotations of different granularity, thereby making it possible to use more abundant heterogeneous training data. At test time, the kinematic spec...

---

*Auto-collected on 2026-06-16.*