Research map / 3D vision / 3D generation
URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Generates simulation-ready URDF articulated objects end-to-end from a single RGB image via part-by-part autoregressive latent diffusion, and validates zero-shot sim-trained manipulation on a real robot via a Real-Follow-Sim digital-twin scheme.
Problem
Recovering executable articulated object models (part geometry plus joint axis, origin, type, and limits) from a single image is ambiguous and usually handled by multi-stage pipelines that depend on retrieval libraries or part segmentation.
Approach
An autoregressive diffusion framework over a TripoSG VecSet latent space that sequentially generates each articulated part with its joint parameters from a shared latent, with an [EoT] token deciding the number of parts, conditioned on DINOv3 image features and a whole-object 3D latent.
Key results
- On the main table, URDF-Anything+ reports whole-object IoU 0.930 and CD 0.009, versus 0.813 IoU and 0.068 CD for the strongest baseline PhysX-Anything.
- Joint axis error 0.129 rad and origin error 0.062 m, versus 0.191 and 0.103 for PhysX-Anything; mean per-part inference time 34.70 s versus 578.38 s for PhysX-Anything.
- Real-world success: Real-Follow-Sim reaches 100% on closing a laptop and 90% on closing a drawer, versus 80%/75% for conventional sim-to-real.
Surprising findings
- Simulation-only training under Real-Follow-Sim matched the real-data-only baseline (100%/90% vs 85%/75%), outperforming it on both tasks.
- Single-layer zero-thickness meshes in PartNet and PhysX-3D caused encoder-decoder reconstruction failures, which was addressed by mesh thickening.
Limitations
- Limited diversity and imbalance of articulated object categories in the training data.
- Does not model mass, friction, or material dynamics, so simulated physics may differ from the real world.
- Real-world evaluation is limited to two tasks with 20 trials each.
Experimental setup
Datasets: PartNet PartNet-Mobility PhysX-3D
Benchmarks: Held-out articulated object test set (PartNet-Mobility/PhysX-3D-derived) Real-world Franka manipulation (close laptop, close drawer)
Metrics: IoU F-score Chamfer Distance joint axis error joint origin error joint limit error inference time per part real-world success rate
Baselines: URDFormer Articulate-Anything SINGAPO PhysX-Anything
Builds on: AutoPartGen TripoSG DiT Diffusion Policy Articulate-Anything
Closest papers by meaning
- URDF-Anything: Constructing Articulated Objects with 3D Multimodal Language Model (NeurIPS 2025 · similarity 0.91)
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 2025 · similarity 0.88)
- Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model (ICLR 2025 · similarity 0.87)
- DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data (NeurIPS 2025 · similarity 0.87)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.87)
- Articulated Kinematics Distillation from Video Diffusion Models (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.