Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Discovering Conceptual Knowledge with Analytic Ontology Templates for Articulated Objects

AAAI 2025 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Imitation and skill learning · 1 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 Analytic Ontology Templates (AOT), parameterized differentiable program descriptions of geometric and kinematic concepts, and AOTNet, a baseline that trains on synthetic AOT-rendered point clouds to discover concepts on articulated objects and interact with them.

Problem

Enabling machine agents to understand and interact with articulated objects from novel categories without real training data, given only raw point clouds of initial and final states.

Approach

Describe geometric and kinematic conceptual ontologies as differentiable analytic templates (structure, parameters, affordances, renderer), train an estimator on synthetic renderings of these templates, and ground affordances analytically to drive interaction.

Key results

• AOTNet achieves 63.7% average success rate on 17 interaction tasks across 653 objects, versus 19.6-46.4% for SAC/TQC RL baselines.
• Human-discovered AOTs reach 97.5% success rate as a reference.
• AOT-state RL improves over raw-point-cloud RL (SAC 34.3 vs 19.6; TQC 39.0 vs 22.6).

Measured effects reported in the paper

Surprising findings

• AOTNet without any real training data outperforms RL baselines that use raw point clouds and some even ground-truth joint and pose information (Raw+JP) for SAC and TQC only partially; Raw+JP TQC (46.4%) exceeds AOTNet.

Limitations

• Only one-gripper interaction tasks on 15 categories; evaluated mainly in simulation with limited real-world demos.
• Simple baseline networks; the authors do not delve into complex designs.
• Composite AOT parameter regression is done at basic-ontology level, so complex composites are harder to estimate directly.

Experimental setup

Datasets: PartNet-Mobility synthetic AOT renderings

Benchmarks: PartNet-Mobility interaction tasks in SAPIEN

Metrics: average success rate (>80% of moving range)

Baselines: SAC TQC Human

Builds on: Point2Mesh Soft Actor-Critic Truncated Quantile Critics

Closest papers by meaning

• ConceptFactory: Facilitate 3D Object Knowledge Annotation with Object Conceptualization (NeurIPS 2024 · similarity 0.85)
• Learning Visually Grounded Domain Ontologies via Embodied Conversation and Explanation (AAAI 2025 · similarity 0.84)
• SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation (NeurIPS 2025 · similarity 0.83)
• OmniEVA: Embodied Versatile Planner via Task-Adaptive 3D-Grounded and Embodiment-aware Reasoning (ICLR 2026 · similarity 0.83)
• Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models (NeurIPS 2026 · similarity 0.83)
• GMAP: Generalized Manipulation of Articulated Objects in Robotic Using Pre-trained Model (AAAI 2025 · similarity 0.83)