CapNav: Benchmarking Vision Language Models on Capability-conditioned Indoor Navigation

CapNav is a benchmark for Vision Language Models evaluating on capability-conditioned navigation reasoning in indoor environments. The benchmark focuses on determining whether an embodied agent with specific physical constraints and abilities can navigate from a start area to a target area within a complex indoor scene.

This repository contains two complementary datasets:

1. The main CapNav Benchmark Dataset, which provides navigation questions paired with scene graph nodes and agent identifiers.
2. The Agent Profiles Dataset, which defines the physical dimensions and capabilities of each agent type.

---

Quickstart

from datasets import load_dataset

capnav = load_dataset("RichardC0216/CapNav", "capnav_v0", split="train")
agents = load_dataset("RichardC0216/CapNav", "agent_profiles", split="train")

print(len(capnav), capnav.column_names)
print(len(agents), agents.column_names)

Dataset Overview

1. CapNav Benchmark Dataset (capnav_v0)

• File: capnav_v0_with_answer.parquet
• Format: Parquet
• Split: train
• Rows: 2,330 (question × agent combinations) This dataset contains natural language navigation questions grounded in indoor scenes. Each question is evaluated under a specific agent profile, enabling systematic analysis of how agent capabilities affect navigation feasibility.

Data Fields

Below one can find the description of each field in the dataset.

• question_id (str)
  Unique identifier for each question–agent pair.
• question (str)
  Natural language navigation question describing whether an agent can move from a start area to a goal area.
• scene_id (str)
  Identifier of the indoor scene (e.g., HM3D00000, MP3D00027).
• scene_type (str)
  High-level category of the scene (e.g., home).
• scene_nodes (list[dict])
  List of scene graph nodes available in the environment.
  Each node contains:
  • node_id (str): Unique identifier of the node
  • name (str): Human-readable node name (e.g., room or area label)
• agent_name (str)
  Name of the agent under which the navigation question should be evaluated. This field links to an entry in the Agent Profiles Dataset.
• answer ((bool))
  Binary ground-truth navigability label (True / False) for the (question, scene, agent) triple.

  Note: answer only provides a binary feasibility signal. For scene graphs, route-level traversability, and detailed ground-truth rationale, please refer to the full annotations in ground_truth/(see below).

---

2. Agent Profiles Dataset

• File: agent_profiles.parquet
• Format: Parquet
• Rows: One per agent type This dataset defines the physical properties and functional capabilities of each agent used in the CapNav benchmark.

Data Fields

• agent_name (str)
  Unique identifier of the agent (e.g., HUMAN, WHEELCHAIR, SWEEPER).
• body_shape (str)
  Abstract geometric representation of the agent’s body (e.g., cylinder, box).
• body_height_m (float)
  Agent height in meters.
• body_width_m (float)
  Agent width in meters.
• body_depth_m (float, optional)
  Agent depth in meters.
  May be null for rotationally symmetric agents.
• max_vertical_cross_height_m (float)
  Maximum vertical obstacle height the agent can cross.
• can_go_up_or_down_stairs (bool)
  Indicates whether the agent is capable of traversing stairs.
• can_operate_elevator (bool)
  Indicates whether the agent can operate and use elevators.
• can_open_the_door (bool)
  Indicates whether the agent can open doors independently.
• description (str)
  Natural language description summarizing the agent’s physical and functional characteristics.

---

Full Ground Truth Annotations (Graphs and Traversability)

This repository also includes a ground_truth/ directory that provides the complete ground-truth annotations used to derive the binary answer labels in the main CapNav benchmark.

While the benchmark dataset exposes only a binary navigability outcome for each (question, scene, agent) triple, the annotations in ground_truth/ contain the underlying structural and traversability information that supports more detailed inspection and analysis.

Specifically, this directory includes:

• Scene graph annotations (ground_truth/graphs/)
  Manually constructed graph representations for each indoor environment, encoding nodes, connectivity, and true spatial structure.
• Agent-conditioned traversability annotations (ground_truth/traverse/)
  Route-level annotations that specify whether and why a path is traversable for a given agent, including agent-specific constraints and failure rationales.

The binary answer field in the benchmark dataset is a distilled signal derived from these annotations. Researchers interested in path validity, failure cases, or the reasons behind infeasible navigation decisions should refer to the files in ground_truth/.

Detailed descriptions of file formats and annotation semantics can be found in:

• ground_truth/README.md

The annotation pipeline, tooling, and quality control procedures are documented in the CapNav GitHub repository:

• https://github.com/makeabilitylab/CapNav
  (see the annotation code and documentation)

---

Intended Use

The CapNav benchmark is intended for:

• Evaluating multimodal or vision-language models on embodied navigation reasoning
• Studying how physical capabilities affect navigation feasibility
• Benchmarking agent-aware reasoning and planning systems The dataset is not intended to provide low-level control signals or precise geometric navigation paths.

---

License

• Dataset: Creative Commons Attribution 4.0 International (CC BY 4.0)

---

Citation

If you find it useful for your research and applications, please cite our paper using this BibTeX:

@article{su2026capnav,
    title={CapNav: Benchmarking Vision Language Models on Capability-conditioned Indoor Navigation},
    author={Su, Xia and Chen, Ruiqi and Liu, Benlin and Ma, Jingwei and Di, Zonglin and Krishna, Ranjay and Froehlich, Jon},
    journal={arXiv preprint arXiv:2602.18424},
    year={2026}
}