Research map / 3D vision / Gaussian splatting

IAAO: Interactive Affordance Learning for Articulated Objects in 3D Environments

CVPR 2025 · 3D vision › Gaussian splatting · approach: Animatable Gaussian human avatars · 2 citations (OpenAlex)

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.

Proposes IAAO, a 3D Gaussian Splatting framework that distills SAM/CLIP/DINOv2 features into hierarchical feature and label fields across two joint states to estimate affordances and articulation parameters for objects and indoor scenes.

Problem

Enabling agents to understand articulated objects and fine-grained functional affordances (e.g., handles, switches) in complex 3D scenes from multi-view images in two joint states, without task-specific networks or assumptions about movable parts, static alignment, or known camera poses.

Approach

Builds per-state 3DGS models with distilled hierarchical CLIP/DINO feature fields and 3D-consistent mask labels, then performs text/point/mask queries on Gaussians, estimates global scene transformation via GeoTransformer and local part motion via 2D-3D matching losses, and fuses the two states.

Key results

• On the PARIS dataset, IAAO attains the best average Axis Ang error (6.32 vs DigitalTwinArt 10.11 on the combined set per Table 1 rows) and Part Motion error 0.81 overall, outperforming Ditto and PARIS.
• On the multi-part dataset (Table 2), IAAO matches or slightly beats DigitalTwinArt in most metrics, e.g., Fridge-m CD-s 0.54 vs 0.61.
• Ablation on multi-part objects shows removing the 2D-3D matching loss raises Axis Ang error from 0.28 to 4.59 and Part Motion from 0.11 to 8.57.

Measured effects reported in the paper

Surprising findings

• Competitor DigitalTwinArt is often on par with or better than IAAO on several multi-part metrics, so the claimed state-of-the-art is not uniform across tables.

Limitations

• Relies on multi-view images from two joint states and per-scene optimization.
• Mesh extraction for evaluation via SuGaR; results on real data limited to few objects.
• Global alignment is coarse and needs refinement; fused scenes may need optional fine-tuning.

Experimental setup

Datasets: PARIS two-part object dataset PartNet-Mobility MultiScan OmniSim

Benchmarks: PARIS two-part object dataset multi-part object dataset OmniSim

Metrics: Axis Ang Err Axis Pos Err Part Motion Err Chamfer-L1 distance (CD-w, CD-s, CD-m)

Baselines: Ditto PARIS PARIS* CSG-reg 3Dseg-reg DigitalTwinArt

Builds on: 3D Gaussian Splatting PARIS Ditto DigitalTwinArt MaskCLIP

Closest papers by meaning

• AffordBot: 3D Fine-grained Embodied Reasoning via Multimodal Large Language Models (NeurIPS 2025 · similarity 0.90)
• Learning 2D Invariant Affordance Knowledge for 3D Affordance Grounding (AAAI 2025 · similarity 0.89)
• Articulate-Anything: Automatic Modeling of Articulated Objects via a Vision-Language Foundation Model (ICLR 2025 · similarity 0.89)
• SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model (CVPR 2025 · similarity 0.89)
• Grounding 3D Object Affordance with Language Instructions, Visual Observations and Interactions (CVPR 2025 · similarity 0.89)
• InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing (CVPR 2025 · similarity 0.88)