IntentionNav Dataset

IntentionNav is a 500-item benchmark for intent-driven object navigation in indoor scenes. Each item describes a target object indirectly through a human intent rather than naming the object category.

Items: 500
Scenes: 176 Kujiale apartment scenes
Target categories: 64
Instruction variants: 4 per item, for 2000 English instructions
Photo format: 1024 x 1024 PNG, rendered in Isaac Sim
Target policy: each item points to one USD object instance whose target category has exactly one physical instance in that scene.

Primary intent-mode counts are: event-script 202, inner-state 167, physical-state 72, and affordance 59.

Layout

README.md
croissant.json
selected_500_intents.jsonl
episodes.jsonl
<scene_id>/
  intents.json
  manifest.json
  photos/
    *.png

Fields Per Item

• selection_id: unique ID, e.g. SEL_001
• scene_id: kujiale_XXXX
• target_category: ground-truth target object category
• target_representative: specific object instance (e.g. air_purifier_0003/Meshes)
• photo: relative path to the image (inside this scene's photos/ dir after reorg)
• formal_en / natural_en / casual_en / emotional_en: 4 English intent variants
• _refine_meta: optional rewrite/judge metadata with target-grounding (tg) and style-distinguishability (sd) scores
• _intent_mode: primary diagnostic mode (EVENT_SCRIPT, INNER_STATE, PHYSICAL_STATE, or AFFORDANCE)
• room / room_type: where the target is located
• see selected_500_intents.jsonl for the full schema

Fixed Navigation Episodes

episodes.jsonl contains the frozen 500-episode navigation specification used for evaluation. Each row is joined to the intent records by selection_id and includes the exact scene_id, target object and position, start position and quaternion, start/target rooms, and geodesic/euclidean distances. The file was frozen before submission and is released unchanged.

Reference Evaluation Resources

The reference_results/ directory additionally provides the reference-agent records, retained episode artifacts, evaluation-code snapshots, and reproducibility metadata used to audit and replay the reported benchmark scores.

The release is organized as follows:

reference_results/
  logs/
  episode_artifacts/
  evaluation_code/
  reproducibility/
  MANIFEST.json
  SHA256SUMS

Archives preserve paths relative to the IntentionNav repository root. MANIFEST.json records archive contents and source revisions, while SHA256SUMS provides release-level integrity checks.

Loading Example

import json

with open("selected_500_intents.jsonl", "r", encoding="utf-8") as f:
    items = [json.loads(line) for line in f if line.strip()]

item = items[0]
print(item["selection_id"], item["scene_id"], item["target_category"])
print(item["natural_en"])

with open("episodes.jsonl", "r", encoding="utf-8") as f:
    episodes = {row["selection_id"]: row for row in map(json.loads, f)}

episode = episodes[item["selection_id"]]
print(episode["start_position"], episode["target_position"])

Image paths are scene-relative. For an item with scene_id == "kujiale_0262" and photo == "surface_photos/21_x.png", the packaged image is under kujiale_0262/photos/21_x.png.

Notes

The dataset is intended for benchmark evaluation of embodied AI agents and VLM-based navigation systems. It does not contain human subjects or personally identifiable information. The current release includes model-generated English intents and automated VLM judge metadata; users should treat the annotations as benchmark labels rather than naturally collected human utterances.

Pinned upstream scene revisions and simulator/environment metadata are included under reference_results/reproducibility/; large third-party scene assets remain hosted by their original providers under their applicable licenses. The 176 scene assets should be downloaded from Eyz/VLNVerse_scene; the exact file list is provided in reference_results/reproducibility/vlnverse_scene_manifest.jsonl.zst.

This peer-review snapshot intentionally omits author identities, affiliations, and identifying project links. The benchmark payload is otherwise unchanged from the fixed pre-submission release.

September shared-category comparison

The versioned system comparison provides the reviewed-v3 inputs (500 tasks, 176 scenes, 65 categories), 500 cached Qwen3.5-4B category predictions, frozen reference/MTU3D code, 2,000 full-set trajectories and 320 development-repeat trajectories. The full set contains 1,500 new runs and 500 verified reused reference runs. The original 64-category reference release above is preserved.

Download system_comparison/20260924/shared_category_20260924.tar.gz, verify its adjacent SHA256SUMS, extract it, then run python recompute.py . inside the extracted folder (requires NumPy and PyYAML). This recomputes all 2,320 trajectory labels and the full-set GSR values without Isaac Sim or hosted API calls. The bundle includes annotation geometry, inspection renders, actual run configurations, source hashes and per-episode tables.

Repeated hosted executions

The versioned hosted repeat study contains 480 full-pipeline executions: 40 uniformly sampled original tasks, four expressions, three repetitions, with Gemini-3.1-Flash-Lite. All 480 episodes generate a fresh initial plan. The bundle includes original inputs, frozen policy and execution wrappers, every trajectory and planner-call receipt, retained images/responses, logs and checksums.

Download hosted_repeats/20260925/hosted_repeats_20260925.tar.gz, verify its adjacent SHA256SUMS, extract it, then run python3 recompute.py . inside the extracted folder (requires NumPy). This verifies the distributed records and reproduces within/across-expression disagreement, paired task/scene intervals and per-repeat SR/All-four without simulator or model calls. An earlier batch that reused initial plans contributes no episode to this release.