GEOPARD: Geometric Pretraining for Articulation Prediction in 3D Shapes

2025-04-03

AI Summary

This paper addresses the problem of joint motion prediction for single-frame static 3D shapes—i.e., inferring kinematic joint constraints between parts without any human-annotated joint labels. The proposed method introduces a geometry-driven self-supervised pretraining paradigm: leveraging physical validity (interpenetration-free and collision-free motion) as a geometric constraint, it designs a geometric search strategy to enable zero-shot discovery and cross-shape generalization of joint constraints. A Transformer-based architecture jointly models part-wise geometric relationships, motion feasibility constraints, and self-supervised contrastive learning. Evaluated on the PartNet-Mobility dataset, the approach achieves state-of-the-art performance, significantly improving both physical plausibility and generalization capability of predicted joint constraints. To the best of our knowledge, this is the first end-to-end framework capable of discovering kinematic joint constraints in fully unlabeled 3D shapes.

Technology Category

• Computer Vision: Motion & Tracking
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We present GEOPARD, a transformer-based architecture for predicting articulation from a single static snapshot of a 3D shape. The key idea of our method is a pretraining strategy that allows our transformer to learn plausible candidate articulations for 3D shapes based on a geometric-driven search without manual articulation annotation. The search automatically discovers physically valid part motions that do not cause detachments or collisions with other shape parts. Our experiments indicate that this geometric pretraining strategy, along with carefully designed choices in our transformer architecture, yields state-of-the-art results in articulation inference in the PartNet-Mobility dataset.

Problem

• Predict articulation from single static 3D shapes
• Learn plausible articulations via geometric pretraining
• Ensure physically valid motions without collisions

Innovation

• Transformer-based architecture for 3D articulation prediction
• Geometric pretraining without manual annotation
• Automatically discovers physically valid part motions