Research map / Large language models / Agents, code and math

OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning

NeurIPS 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents

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 OntoPlan, an ontology-grounded world model aligned with PDDL plus an agentic framework with Scene query and PDDL plan tools, achieving 0.89 task success on 150 indoor robot tasks with 18.1k tokens per task.

Problem

LLM-based robot task planning degrades on long-horizon tasks in large scenes: text-serialized spatial graphs cause spatial reasoning errors and token cost grows with scene size, and LLM-generated action sequences often violate preconditions and goals.

Approach

Keep scene knowledge in an OWL ontology aligned one-to-one with the PDDL domain, let agents query it selectively via SPARQL/embedding search, and delegate executable plan synthesis to Fast Downward with ordered subgoals and state write-back, rather than generating action sequences directly with the LLM.

Key results

• OntoPlan achieves 0.89 average task success vs 0.27 for SayPlan and 0.13 for DELTA on general tasks.
• OntoPlan uses 18.1k total tokens per task on average vs 216.0k (SayPlan) and 100.6k (DELTA).
• Removing the PDDL plan tool drops task success from 0.89 to 0.31.

Surprising findings

• Removing the Scene query tool leaves success nearly unchanged (0.87) while token use grows about 4x from Small to Large scale.

Limitations

• Assumes fully observable, fully instantiated symbolic world model; no raw-sensor perception evaluated.
• Ontology-PDDL alignment adds modeling cost; new semantics or actions require coordinated schema extensions.
• Evaluated on a small set of hand-written instructions (150 general, 10+10 diagnostic) in five environments.

Experimental setup

Datasets: 3D Scene Graph dataset (Klickitat, Lakeville, Lindenwood, Marstons, Muleshoe)

Benchmarks: 150 general tasks across 5 environments and 3 scales 10 ambiguous tasks 10 infeasible tasks

Metrics: plan returned plan executable goal reached task success input tokens per call total tokens per task

Baselines: SayPlan DELTA w/o Scene query w/o PDDL plan

Builds on: SayPlan DELTA LLM-modulo Fast Downward PDDL

Closest papers by meaning

• R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.89)
• Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.89)
• GNN-Transformer Task Planning Enhanced with Semantic-Driven Data Augmentation (AAAI 2025 · similarity 0.88)
• ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making (NeurIPS 2026 · similarity 0.87)
• Graph-enhanced Large Language Models in Asynchronous Plan Reasoning (ICML 2024 · similarity 0.87)
• ClevrSkills: Compositional Language And Visual Reasoning in Robotics (NeurIPS 2024 · similarity 0.87)