Skip to content

Articulation in Motion: Prior-Free Part Mobility Analysis for Articulated Objects

Conference: ICLR 2026 arXiv: 2603.02910 Project Page: AiM Area: Other Keywords: articulated objects, Gaussian splatting, part segmentation, joint estimation, sequential RANSAC, prior-free, interaction video

TL;DR

This paper proposes AiM (Articulation in Motion), a framework that reconstructs articulated objects from interaction videos and initial-state scans without requiring prior knowledge of the number of parts. It achieves dynamic-static decoupling via a dual-Gaussian representation (Static GS + Deformable GS), combines sequential RANSAC for prior-free part segmentation and joint estimation, and incorporates an SDMD module to handle newly exposed static regions. On complex 6-part objects (Storage), AiM achieves 79.34% mean IoU, substantially outperforming the prior-dependent ArtGS (52.23%).

Background & Motivation

Core demand for articulated object understanding: Robot manipulation, AR/VR, and embodied intelligence all require understanding the part structure and joint parameters of articulated objects (e.g., drawer cabinets, doors, laptops).

Prior dependency of existing methods: Methods such as DTA and ArtGS require the number of parts to be specified in advance, which is typically unknown in real-world scenarios; an incorrect specification leads to severe segmentation failures.

Challenge of dynamic-static decoupling: During interaction, some parts move while others remain static; however, the displacement of moving parts exposes previously occluded static regions, which conventional methods struggle to handle.

Limitation of single representations: Purely static or purely dynamic 3D Gaussian representations cannot simultaneously accommodate the mixed nature of fixed and moving parts in articulated objects.

Diversity of joint types: Articulated objects contain multiple joint types including revolute and prismatic joints, necessitating a unified prior-free estimation approach.

Practicality of video input: Recovering articulation information from a single interaction video is more practical and natural than methods requiring multi-view static scans.

Method

Overall Architecture

AiM takes as input an interaction video of a human manipulating an articulated object and a 3D scan of the object in its initial (static) state, and outputs part segmentation, joint parameters, and a complete articulated object reconstruction. The pipeline consists of three stages: dual-Gaussian dynamic-static decoupling → sequential RANSAC part discovery → joint parameter estimation.

Key Designs

  1. Dual-Gaussian Representation

  2. Function: Maintains two sets of 3D Gaussians — Static GS representing invariant background and stationary parts, and Deformable GS representing moving parts.

  3. Mechanism: Gradient signals from pixel-level rendering losses automatically assign Gaussians to static or dynamic sets; Static GS remains fixed while Deformable GS learns per-frame deformation fields.

  4. Design Motivation: Explicit dynamic-static separation prevents moving parts from corrupting static geometry, and allows subsequent part segmentation to focus exclusively on dynamic Gaussians.

  5. Sequential RANSAC Part Segmentation

  6. Function: Automatically discovers parts from the motion trajectories of dynamic Gaussians without presetting the number of parts.

  7. Mechanism: Fits rigid body motion to the deformation trajectories of all dynamic Gaussians; the largest consensus set corresponds to one part. That part is then removed and the process iterates over the remaining Gaussians until the residual falls below a threshold.

  8. Design Motivation: RANSAC is naturally suited to the setting of "an unknown number of mixed rigid body motions"; sequential execution ensures parts are discovered in descending order of size.

  9. SDMD Module (Static Dynamic Merging with Discovery)

  10. Function: Handles static regions newly exposed after moving parts are displaced (e.g., the interior walls of a cabinet revealed when a drawer is opened).

  11. Mechanism: Detects discrepancy regions between rendered and real images, initializes new Static Gaussians at those locations, and merges them with the existing Static GS.
  12. Design Motivation: Conventional methods cannot handle static geometry that is initially invisible but later becomes observable; SDMD fills this critical gap.

Key Experimental Results

Main Results

Method Part Prior Mean IoU (%) Revolute JE (°) Prismatic JE (mm)
DTA Required 71.45 8.32 12.7
ArtGS Required 76.99 5.61 8.9
AiM (Ours) Not Required 80.21 4.23 7.1

Ablation Study

Component Mean IoU (%) Note
Full AiM 80.21 Complete method
w/o SDMD 74.85 Newly exposed regions incorrectly assigned
Single GS (no decoupling) 68.32 Moving parts corrupt static reconstruction
K-means instead of RANSAC 72.56 Requires preset K and is sensitive to noise
ArtGS with ground-truth part count 76.99 Still underperforms AiM even with correct prior

Key Findings

  1. Prior-free surpasses prior-dependent: AiM achieves 80.21% mean IoU without part-count priors, exceeding prior-dependent ArtGS (76.99%), demonstrating that adaptive discovery is more robust than fixed assumptions.
  2. Decisive advantage on complex objects: On the 6-part Storage object, AiM (79.34%) vs. ArtGS (52.23%) shows a gap of 27%; ArtGS degrades sharply as the number of parts increases.
  3. SDMD is indispensable: Removing SDMD causes a 5.36% IoU drop, confirming the importance of handling newly exposed regions.
  4. Dynamic-static decoupling is foundational: The single-GS variant underperforms the full method by nearly 12%, establishing the dual-Gaussian design as the cornerstone of success.

Highlights & Insights

  1. Complete elimination of priors: AiM is the first method to achieve part segmentation and joint estimation for articulated objects without requiring prior knowledge of the number of parts, better matching real-world application demands.
  2. Elegant dual-Gaussian decoupling: Embedding dynamic-static separation into the 3DGS representation simultaneously benefits reconstruction quality and downstream analysis.
  3. Practical innovation of SDMD: Addresses the progressive exposure of previously occluded static regions — a critical yet often overlooked detail in articulated object understanding.
  4. Natural fit of sequential RANSAC: Cleverly exploits the iterative stripping property of RANSAC to achieve adaptive part-count discovery.
  5. Overwhelming advantage on complex objects: The 27% improvement on 6-part scenes demonstrates the scalability of the approach.

Limitations & Future Work

  1. Single-interaction assumption: The current method requires that all parts be actuated within the video; parts that are not manipulated cannot be discovered.
  2. Rigid body motion assumption: Sequential RANSAC assumes each part undergoes rigid body motion and cannot handle flexible hinges or elastic deformations.
  3. Computational cost: The combination of dual-Gaussian representation and sequential RANSAC incurs substantial computational overhead, precluding real-time operation.
  4. Dependency on video quality: Low-quality videos with severe motion blur or occlusion may lead to inaccurate dynamic Gaussian estimation.
  • Articulated object reconstruction: Gaussian splatting-based methods including DTA (Liu et al., 2024) and ArtGS (Huang et al., 2024).
  • 3D Gaussian Splatting: 3DGS (Kerbl et al., 2023), Dynamic 3DGS (Luiten et al., 2024).
  • Part segmentation: Supervised methods such as PartNet (Mo et al., 2019); unsupervised methods such as SAM3D.
  • RANSAC: The classic framework of Fischler & Bolles (1981); application of sequential RANSAC to multi-model fitting.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Prior-free part discovery + dual-Gaussian decoupling + SDMD are all novel designs.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Validated across multiple object categories with comprehensive ablations.
  • Writing Quality: ⭐⭐⭐⭐ Method pipeline is clearly presented; experimental results are detailed.
  • Value: ⭐⭐⭐⭐⭐ Prior-free articulated object understanding has significant practical value for robotics and embodied AI.