Alessandro Leanza

LOCAL LLMs / GRAPH RETRIEVAL / ROBOT PLANNING

Graph-Based RAG for Robot Planning

Graph-Based Retrieval-Augmented Planning for Robots with Provenance-Aware Knowledge

Accepted for publication in the DSA ISC 2026 proceedings (Springer, LNCS, vol. 16621).

A fully local task-planning framework that turns object-associated text into a persistent knowledge graph with provenance. It retrieves a compact, scene- and task-relevant subgraph, then uses that evidence to guide a local LLM toward grounded JSON subgoals.

Authors: Alessandro Leanza · Loris Roveda · Blerina Spahiu

CodeBibTeX
Object-associated text is converted into evidence-linked facts, retrieved as a task-relevant subgraph, and passed to a constrained local LLM planner
Provenance-aware graph planning architecture

The problem

LLM planners can reference objects or actions that are unavailable, or make unsupported assumptions about materials, hazards, and affordances. A transient prompt does not provide a reusable record of what the system knows or where that knowledge came from.

This project builds an inspectable, object-centric memory from scene-associated evidence and retrieves only the context needed for the current instruction.

Approach

  1. Build

    Object-centric knowledge memory

    Chunk object-associated documents and extract schema-constrained facts with a local LLM. Store each fact in Neo4j with confidence, source metadata, and links to supporting evidence.

  2. Retrieve

    Scene- and task-conditioned context

    Retrieve facts for the perceived objects and select a compact, diverse subgraph under a prompt budget, balancing source reliability, confidence, and instruction relevance.

  3. Plan

    Grounded JSON subgoals

    Condition a local LLM on the instruction, scene entities, and selected evidence. Produce ordered subgoals using a restricted action-template library for downstream skills.

Key results

The manuscript compares task decomposition across six categories from the ConceptBot task suite. The selected results below are reported symbolic planning outcomes, not physical robot task-success rates.

86%

Materials

GraphRAG

ConceptBot 70% · SayCan 20% · ConceptBot w/o ConceptNet 51%

96%

Toxicity

GraphRAG

ConceptBot 86% · SayCan 36% · ConceptBot w/o ConceptNet 56%

85%

Implicit tasks

GraphRAG

ConceptBot 87% · SayCan 31% · ConceptBot w/o ConceptNet 64%

75%

Risk-aware tasks

GraphRAG

ConceptBot 76% · SayCan 15% · ConceptBot w/o ConceptNet 58%

GraphRAG improves on full ConceptBot in Materials and Toxicity; full ConceptBot remains slightly stronger on Implicit and Risk-Aware tasks. These results do not establish grasping, collision avoidance, placement stability, or controller robustness.

Scope and limitations

The framework operates at the semantic/task-decomposition layer. Restricted action templates reduce unsupported outputs, but do not replace reachability, graspability, geometric, or interaction-risk checks in downstream robot skills.

Knowledge coverage matters: unresolved object names, missing sources, and broad material or toxicity descriptions can still lead to incorrect plans. Provenance makes those dependencies inspectable; it does not certify the facts or guarantee safe execution.

Technical snapshot

Graph backend
Neo4j · Object, Fact, and Chunk nodes
Knowledge sources
Object-associated text; Wikipedia in the prototype
Extraction
Qwen3-Coder:30B
Planning
Qwen3:32B
Relevance embeddings
all-MiniLM-L6-v2
Local model serving
Ollama
Output
JSON subgoals with restricted action templates
Provenance
Source metadata, confidence, and evidence-chunk links
Implementation
Python · Docker

Publication

Graph-Based Retrieval-Augmented Planning for Robots with Provenance-Aware Knowledge

Alessandro Leanza, Loris Roveda, Blerina Spahiu

Accepted for publication in the DSA ISC 2026 proceedings (Springer, LNCS, vol. 16621).

Paper ID 25. The proceedings are forthcoming; final bibliographic metadata and DOI will be added when available. No arXiv version is currently linked.

Code

BibTeX — provisional citation

This is a provisional author/title citation for an accepted, forthcoming paper. It is not the final Springer citation; no DOI, page range, or formal volume title is asserted.

@unpublished{leanza2026graphbased,
  title={Graph-Based Retrieval-Augmented Planning for Robots with Provenance-Aware Knowledge},
  author={Leanza, Alessandro and Roveda, Loris and Spahiu, Blerina},
  year={2026},
  note={Accepted for publication in the DSA ISC 2026 proceedings, Springer, Lecture Notes in Computer Science, vol. 16621, Paper ID 25. Final publication metadata pending.}
}