Reasoning and Planning

VIRF

VIRF combines scene knowledge graphs, ontology-based safety rules, and iterative plan verification so embodied agents can generate safer task plans in household environments.

Tool Introduction

Core parameters, trigger timing, and visual before/after demo references.

Short Explanation

Use VIRF when a plan must satisfy explicit safety rules instead of relying on unconstrained generative reasoning alone.

InputUser task + scene knowledge base

OutputSafety status, verified plan, knowledge-base updates

Trigger TimingTriggered on demand after the required input files and configuration are prepared.

RuntimePython / JSON knowledge base / optional ontology backend

BeforeUser task + scene knowledge base

Prepare the scene, image, video, sensor stream, prompt, or configuration expected by the original project.

AfterSafety status, verified plan, knowledge-base updates

Read the produced visualization, prediction, map, trajectory, mask, grasp pose, or other documented artifact.

Preset Example

A quick-run style example for the documentation page.

Inputtools/virf/examples/test_input.json

PromptCusr: Heat the soup in the pot

ExpectedA SAFE or unsafe status plus a revised multi-step plan that respects the rules encoded in the knowledge base.

Parameters And Output

Readable controls and the meaning of each returned artifact.

Parameter Explanation

Cusrtext

User task description to plan and verify.

Kpathtools/virf/examples/test_input.json

Knowledge base payload containing ABox entities and TBox safety rules.

max_attemptsnumber3

Maximum number of refinement attempts used by the main verification loop.

Output Explanation

status

Safety result such as SAFE.

safe_plan

Verified task plan that satisfies the available rules.

knowledge_graph

Structured world state consumed by the verifier.

How To Use

Official resources, deployment steps, academic context, citation, and source-reported benchmark numbers.

Resources

GitHubhttps://github.com/Sn0wm1an/VIRF Code Downloadhttps://github.com/Sn0wm1an/VIRF/archive/refs/heads/master.zip Paperhttps://arxiv.org/abs/2602.08373

Deployment Notes

Clone the official VIRF repository and install the Python dependencies required by the KG, ontology creator, and agent bench modules.
Prepare an input JSON containing the user command plus the scene knowledge base under `tools/virf/examples/`.
Run `python main.py` to execute the simplified bundled demo or call the agent bench modules for fuller evaluation.
Use the resulting `output.json` to inspect whether VIRF marked the plan safe and how it revised the action sequence.

Relative Path Example

# Relative-path local entry for the VIRF deployment
cd tools/virf
python main.py

# The bundled demo reads tools/virf/examples/test_input.json and writes tools/virf/results/output.json.

Expected Result Shape

{
  "tool": "virf",
  "status": "ok",
  "results": [
    {
      "label": "Safety-verified task reasoning",
      "score": 0.87,
      "output": "Safety status, verified plan, knowledge-base updates"
    }
  ],
  "timing": {
    "runtime": "The deployment README and local demo do not provide a source-reported latency number.",
    "device": "documented in source benchmark when available"
  },
  "artifacts": {
    "visualization": "tools/virf/runs/visualization.png",
    "raw_predictions": "tools/virf/runs/predictions.json"
  }
}

Paper figure

Academic Info

Paper identity and contribution summary.

TitleGrounding Generative Planners in Verifiable Logic: A Hybrid Architecture for Trustworthy Embodied AI

AuthorsFeiyu Wu, Xu Zheng, Yue Qu, Zhuocheng Wang, Zicheng Feng, Hui Li

VenueICLR 2026

ContributionGrounds generative planners in verifiable logic so unsafe plans can be filtered or revised before execution in embodied tasks.

Citation

@misc{virf2026,
  title={Grounding Generative Planners in Verifiable Logic: A Hybrid Architecture for Trustworthy Embodied AI},
  author={Feiyu Wu and Xu Zheng and Yue Qu and Zhuocheng Wang and Zicheng Feng and Hui Li},
  year={2026},
  note={ICLR 2026},
  url={https://arxiv.org/abs/2602.08373}
}

Benchmark

Only compact, source-reported numbers are shown here.

Dataset	Metric	Value	Runtime	Source
SafeAgentBench	HAR / GCR / Avg correction iterations	0.0% / 77.3% / 1.1	Source paper result	ICLR 2026 paper

Artifacts

Official paper, repository README, bundled example JSON, simplified main loop, and output plan JSON.

Demo Images

Visual references from the original tool. Click any image to inspect the original size.