MultINav-Bench

A benchmark for interaction-aware navigation: navigation tasks in which doors, sliding doors, gate-like barriers, fridges, cabinets, and drawers change the structure of the navigable space.

MultINav-Bench evaluates whether a policy can decide when an interaction is worth performing, not only how to perform it. Each episode attaches interactive objects, their state (unknown / closed / open), and the resulting state transition to the scene topology, so a planner can exploit how interaction changes reachability, visibility, and path cost.

The benchmark is defined over MolmoSpaces scenes from procthor-10k (val split) with the rby1 mobile manipulator, and uses the interactive_nav_v3 episode schema.

Domains

Domain	Episodes	Interaction type	Effect on the navigation graph
channel	1000	door, sliding door, gate-like barrier, movable blocker	changes connectivity / reachability
container	1000	fridge, cabinet, drawer	changes target visibility and accessibility
mixed	1000	channel then container	chained interaction: one interaction enables the next
all	3000

Files

v1.2/
├── benchmark/
│   ├── benchmark.json     # all 3000 episodes (list of episodes)
│   ├── channel.json       # 1000 channel episodes
│   ├── container.json     # 1000 container episodes
│   └── mixed.json         # 1000 mixed episodes
├── analysis/              # QC, distribution, and paper-facing statistics
│   ├── dataset_qc_report.md
│   ├── distribution_summary.json
│   ├── episodes_flat.{csv,jsonl}
│   ├── paper_data/*.csv
│   └── figures/*.{pdf,png}
├── provenance/            # resolved collection config and repair manifests
├── checksums.sha256
└── README.md

Verify integrity from inside v1.2/:

sha256sum -c checksums.sha256

Episode format

Each episode follows the interactive_nav_episode.schema.json schema (interactive_nav_v3). Top-level fields:

Field	Meaning
house_index, scene_dataset, data_split	source scene (procthor-10k, val)
robot	rby1 with initial qpos
cameras, img_resolution	camera rig and resolution (640x480)
scene_modifications	scene edits applied before the episode
task	nav_to_obj target, thresholds, horizon
language	instruction type, disclosure, referral expressions
interactive_nav	the interactive-navigation ground truth (see below)
source	upstream provenance of the underlying NavToObj episode

interactive_nav contains:

Field	Meaning
schema_version	interactive_nav_v3
case_id	globally unique episode id
interaction_domains	["channel"], ["container"], or ["channel", "container"]
interaction_requirement	required (interaction needed for success) or unnecessary
target	target category / instance and grounding
success_criteria	distance + head-camera visibility conditions
initial_state.interaction_states	per-joint initial state and semantic_state
interactions	interaction objects, joints, and state transitions
oracle_plan / oracle_plans	reference interaction-then-navigation plan

Statistics

From analysis/dataset_qc_report.md and analysis/dataset_health_report.json:

• 3000 episodes, 3000 unique case_id, 0 duplicate and 0 semantic issue.
• Interaction required in 2654 / 3000 episodes (88.5%); channel 65.4%, container and mixed 100%.
• Effect labels: reveal_target_object 2000, restore_reachability 1654, enable_interaction 8.
• Visibility gain: container 100%, mixed 100%, channel 0% (by construction).
• Scenes: 639 houses total (channel 253, container 569, mixed 369).
• Reference GT path length (mean / median, m): channel 8.70 / 8.12, container 7.99 / 5.93, mixed 13.58 / 11.50, all 10.09 / 8.69.
• Target categories: 31 across all domains; container categories: 2 (fridge, cabinet/drawer).

Evaluation

MultINav-Bench is evaluated with the MolmoSpaces nav_to_obj success criterion (distance threshold plus head-camera visibility). The main reported metrics are:

Metric	Meaning
SR	final task success rate
SPL	success-weighted path efficiency (0 on failure)
Interaction Success Rate	necessary interactions completed / necessary interactions, averaged over episodes that need interaction
Interaction Precision	completed interactions of the target category / all interaction attempts
Total Cost	min((L_exec + lambda * A + mu * E) / B, 1) on success, 1 on failure; defaults lambda=0.3, mu=1, B=30

Results are reported per split (all, channel, container, mixed, no-interaction). no-interaction episodes measure interaction restraint, so their interaction success rate may be reported as N/A. See the evaluation protocol for full definitions.

Usage

1. Install the evaluator from MultINav and download the procthor-10k scenes and molmospaces-bench-v2 resources:

pip install -e ".[molmospaces]"
hf download Piqiuni/MultINav-Bench --repo-type dataset --local-dir benchmarks/data

2. Point the evaluator at the benchmark file:

multinav-eval --benchmark benchmarks/data/v1.2/benchmark/benchmark.json --policy <name>

This card is tracked at benchmarks/dataset_card.md in that repository, next to the episode schema it describes.

Provenance

Episodes are derived from the upstream MolmoSpaces NavToObjProcthor10kBench_20260112 benchmark (val split). The source field records the upstream episode that each sample was built from and is provenance-only: the released episodes are self-contained and do not require the referenced upstream H5 files. The provenance/ directory records the resolved collection config and the repair manifests used to build this release.

This release is interactive-nav-v3-procthor10k-val-release-v1.2.

License

Released under CC BY 4.0, consistent with the MolmoSpaces data subsets it is derived from. Objaverse-derived content in the underlying scenes is licensed under ODC-BY 1.0.

Citation

@article{kim2026molmospaces,
  title={MolmoSpaces: A Large-Scale Open Ecosystem for Robot Navigation and Manipulation},
  author={Kim, Yejin and Pumacay, Wilbert and Rayyan, Omar and Argus, Max and Han, Winson and VanderBilt, Eli and Salvador, Jordi and Deshpande, Abhay and Hendrix, Rose and Jauhri, Snehal and others},
  journal={arXiv preprint arXiv:2602.11337},
  year={2026}
}

The MultINav paper citation will be added upon publication.