OntoPlan

From an instruction to an executable plan

• LLM agents plan over an ontology-grounded world model whose classes and properties map one-to-one to a PDDL domain, so grounding, goals, and planning share one vocabulary.
• They query only the task-relevant part of the scene and hand formalized goals and constraints to a symbolic planner, so returned plans are executable.
• On 150 long-horizon tasks in five buildings at three scales: 0.89 task success vs. 0.27 for the strongest baseline, with 5.6× fewer tokens, and cost stays nearly flat as scenes grow 10–23×.

The Flow Orchestrator routes grounding, clarification, task formalization, and planning across specialized agents, while the Scene query and PDDL plan tools operate over a shared ontology-grounded world model.

Higher success at a fraction of the token cost

World model. Ask supports scene inspection and planner-state extraction; tell updates dynamic facts from symbolic observations and action effects, and GraphDB materializes the OWL 2 RL entailments. The Marstons Medium scene has 2,908 asserted triples and 19,394 after reasoning.

One vocabulary from the world model to the planner

:isInsideOf a owl:ObjectProperty ,
                owl:TransitiveProperty ;
    rdfs:domain :Artifact ;
    rdfs:range  :Artifact .

# dynamic layer fact
:toothbrush_213 :isInsideOf :storage_cabinet_55 .

(:types ... Artifact)

(:predicates
  (isInsideOf ?x - Artifact ?c - Artifact)
  ...)

; extracted initial state
(isInsideOf toothbrush_213 storage_cabinet_55)

Watch OntoPlan solve a task

Four runs from the evaluation, replayed step by step: the agents' messages as logged, the conditions and subgoals they produced, and the plan executed on the Klickitat scene graph. Nothing here is regenerated; every message and action comes from the run's result file.

The purple marker is the robot and its trail is the executed path; orange squares are objects being carried or put down, and a red cross marks a place an always condition forbids. In the first run the Flow Orchestrator asks which couch to use and receives the fixed automatic reply used in the non-interactive evaluation.

Success holds and token cost stays flat as scenes grow

From Small to Large, object counts grow from 50–121 to 954–1,269 per environment. OntoPlan keeps the highest rate on every metric at every scale, while SayPlan shows heavier token tails and DELTA remains far more token-intensive.

Ablation: what each tool contributes

Without Scene query, the full scene is serialized into the prompt. Without PDDL plan, the LLM generates action sequences directly, given the same PDDL domain. Tokens are ×10³; L/S is the Large-to-Small ratio.

Selective scene access drives token scalability; the integrated planning path drives task reliability.

Asking before acting

Two qualitative subsets of 10 instructions each in Klickitat (Medium). When an instruction leaves the target open, OntoPlan asks the user; when a request cannot be satisfied, it reports what is missing instead of committing to a wrong plan.

A multi-turn session over one world model

A question, a task, and a follow-up in one session. Expand a tool call to see its arguments and result. Plans run on the shared world model, and each turn ends with the facts it changed, so the next turn starts from the updated state. Recorded with the released code; the paper's figure shows an earlier run of the same session, so details such as where the mug is put down differ.

Dashed blue lines are paths returned by find_path; the purple trail is the executed plan. In Turn 3 the robot starts in the living room where Turn 2 left it, and the bread is still in the switched-on oven.

Five buildings, three scene scales

BibTeX

@inproceedings{nam2026ontoplan,
  title     = {{OntoPlan}: An Ontology-Grounded Scene Representation and
               Agentic Framework for Scalable Robot Task Planning},
  author    = {Nam, Hyeongwoo and Cho, Woongje and Kim, Juwon and Choi, Jongeun},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}