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

A new model called Instruct-Particulate can reconstruct articulated 3D objects with user-specified kinematic control, addressing a long-standing data-scarcity problem in computer vision, according to research posted to the arXiv preprint server [1]. The model accepts a 3D mesh and 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 [2]. The specification disambiguates the task and allows the model to target annotations of different granularity, which the authors say makes it possible to use more abundant heterogeneous training data [2]. At test time, the kinematic specification can be obtained automatically from large-scale vision-language models, so the approach can be applied to any input mesh [2]. To train the system at scale, the researchers constructed a heterogeneous dataset of more than 150,000 articulated 3D objects [2]. They extended existing publicly available collections by partially labeling other 3D models — both monolithic and pre-decomposed — with kinematic labels generated by vision-language models [2]. Experiments showed that the model generalizes better across categories and to AI-generated meshes, and it enables articulated asset reconstruction from real-world images when paired with image-to-3D models [2]. The paper was submitted on 12 June 2026 to the computer-vision section of arXiv, the open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields [6]. As of November 2024, arXiv was receiving about 24,000 new articles per month and had surpassed two million total articles at the end of 2021 [6]. The repository is not peer-reviewed; papers are approved after moderation [6]. The work appears alongside experimental community tools hosted under arXivLabs, a framework launched in 2020 that allows third-party collaborators to build features on top of arXiv’s article pages [4]. “Members of our community want to contribute tools that enhance the arXiv experience, and we value that kind of community engagement,” said Eleonora Presani, arXiv’s executive director, at the launch [4]. Current Labs integrations include the Bibliographic Explorer for citation-tree navigation, the CORE Recommender for surfacing related open-access papers, and links to code repositories via Papers with Code [5]. arXiv has stated that Labs collaborators receive only minimal, anonymized user data and are prohibited from using it for any purpose beyond the correct functioning of the feature [4]. The Instruct-Particulate research targets applications in animation, gaming, and robotic simulations, where reconstructing articulated 3D objects is a foundational task [2].