Datasets:
license: cc-by-4.0
pretty_name: EmbodiedNav-Bench
language:
- en
task_categories:
- visual-question-answering
- reinforcement-learning
tags:
- embodied-ai
- embodied-navigation
- urban-airspace
- drone-navigation
- multimodal-reasoning
- spatial-reasoning
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train-00000-of-00001.parquet
How Far Are Large Multimodal Models from Human-Level Spatial Action? A Benchmark for Goal-Oriented Embodied Navigation in Urban Airspace
Abstract
Large multimodal models (LMMs) show strong visual-linguistic reasoning but their capacity for spatial decision-making and action remains unclear. In this work, we investigate whether LMMs can achieve embodied spatial action like human through a challenging scenario: goal-oriented navigation in urban 3D spaces. We first spend over 500 hours constructing a dataset comprising 5,037 high-quality goal-oriented navigation samples, with an emphasis on 3D vertical actions and rich urban semantic information. Then, we comprehensively assess 17 representative models, including non-reasoning LMMs, reasoning LMMs, agent-based methods, and vision-language-action models. Experiments show that current LMMs exhibit emerging action capabilities, yet remain far from human-level performance. Furthermore, we reveal an intriguing phenomenon: navigation errors do not accumulate linearly but instead diverge rapidly from the destination after a critical decision bifurcation. The limitations of LMMs are investigated by analyzing their behavior at these critical decision bifurcations. Finally, we experimentally explore four promising directions for improvement: geometric perception, cross-view understanding, spatial imagination, and long-term memory.
Dataset Overview
EmbodiedNav-Bench is a goal-oriented embodied navigation benchmark for evaluating how large multimodal models act in urban 3D airspace. The released sample set contains 300 human-collected trajectories with natural-language goals, drone start poses, target positions, and ground-truth 3D paths. The original evaluation data is provided as dataset/navi_data.pkl, and a Parquet conversion is provided at data/train-00000-of-00001.parquet for the Hugging Face Dataset Viewer table.
Navigation Example
| Example 1 | Example 2 | Example 3 |
|---|---|---|
| Goal: Nearby bus stop | Goal: The fresh food shop in the building below | Goal: The balcony on the 20th floor of the building on the right |
 |
 |
 |
Note: The videos above demonstrate goal-oriented embodied navigation examples in urban airspace. Given linguistic instructions, the task evaluates the ability to progressively act based on continuous embodied observations to approach the goal location.
Dataset Statistics
Key Statistics:
- Total Trajectories: 5,037 high-quality goal-oriented navigation trajectories
- Data Collection: Over 500 hours of human-controlled data collection
- Average Trajectory Length: ~203.4 meters
- Annotators: 10 volunteers (5 for case creation, 5 experienced drone pilots with 100+ hours flight experience)
- Action Types:
- Horizontal movement (move-forth, move-left, move-right, move-back)
- Vertical movement (move-up, move-down)
- Rotation/view Change (turn-left, turn-right，adjust-camera-gimbal-upwards, adjust-camera-gimbal-downwards)
- Trajectory Distribution: Pay more attention to vertical movement
Dataset Construction and Statistical Visualization:
Figure: a. Dataset Construction Pipeline. b. The length distribution of navigation trajectories. c. Proportion of various types of actions. d. The relative position of trajectories to the origin. e. Word cloud of goal instructions.
Environment Setup and Simulator Deployment
This project references EmbodiedCity for the urban simulation environment.
1. Download the simulator
- Offline simulator download (official): EmbodiedCity-Simulator on HuggingFace
- Download and extract the simulator package, then launch the provided executable (
.exe) and keep it running before evaluation.
2. Create the Python environment
Use one of the following ways:
conda create -n EmbodiedCity python=3.10 -y
conda activate EmbodiedCity
pip install airsim openai opencv-python numpy pandas
If you are using the simulator package's built-in environment files:
conda env create -n EmbodiedCity -f environment.yml
conda activate EmbodiedCity
3. Dataset release
All paths below are relative to the project root.
We are currently open-sourcing 300 trajectories as public examples:
dataset/navi_data.pkldataset/navi_data_preview.json(human-readable JSON preview)data/train-00000-of-00001.parquet(Hugging Face Dataset Viewer table split)
dataset/navi_data.pkl is the canonical dataset file for evaluation.
3.1 navi_data.pkl field schema
Each sample in dataset/navi_data.pkl is a Python dict with the following fields:
| Field | Type | Description |
|---|---|---|
folder |
str |
Scene folder identifier |
start_pos |
float[3] |
Initial drone world position (x, y, z) |
start_rot |
float[3] |
Initial drone orientation (roll, pitch, yaw) in radians |
start_ang |
float |
Initial camera gimbal angle (degrees) |
task_desc |
str |
Natural-language navigation instruction |
target_pos |
float[3] |
Target world position (x, y, z) |
gt_traj |
float[N,3] |
Ground-truth trajectory points |
gt_traj_len |
float |
Ground-truth trajectory length |
3.2 Example view for humans
To make inspection easier without loading PKL directly, we provide:
dataset/navi_data_preview.json
This JSON contains:
- field descriptions
- total sample count
- preview of the first few samples (including
gt_trajpartial points)
Example item (simplified):
{
"sample_index": 0,
"folder": "0",
"task_desc": "the entrance of the red building on the left front",
"start_pos": [6589.18164, -4162.23877, -36.2995872],
"start_rot": [0.0, 0.0, 3.14159251],
"start_ang": 0.0,
"target_pos": [6390.7041, -4154.58545, -6.29958725],
"gt_traj_len": 229.99981973603806,
"gt_traj_num_points": 28,
"gt_traj_preview_first5": [
[6589.18164, -4162.23877, -36.2995872],
[6579.18164, -4162.23877, -36.2995872],
[6569.18164, -4162.23877, -36.2995872],
[6559.18164, -4162.23877, -36.2995872],
[6549.18164, -4162.23877, -36.2995872]
]
}
3.3 Hugging Face Dataset Viewer table
The train split is stored as data/train-00000-of-00001.parquet so the dataset can be inspected directly in the Hugging Face Table view. Each table row corresponds to one navigation trajectory and includes flattened coordinate columns (start_x, target_x, etc.) together with the original structured fields (start_pos, start_rot, target_pos, and gt_traj).
4. How to test your own model
To evaluate your model, modify the Agent logic in embodied_vln.py, mainly in the ActionGen class:
ActionGen.query(...): replace prompt design / model API call / decision logic.- Keep output command format compatible with
parse_llm_action(...)(one command per step). - Supported commands include:
move_forth,move_back,move_left,move_right,move_up,move_down,turn_left,turn_right,angle_up,angle_down.
Then run:
python embodied_vln.py
Example: connect other API models
Use the API placeholder pattern in embodied_vln.py as a template for plugging in your own model service.
Current placeholders (in embodied_vln.py) are:
AZURE_OPENAI_MODELAZURE_OPENAI_API_KEYAZURE_OPENAI_ENDPOINTAZURE_OPENAI_API_VERSION(optional, default:2024-07-01-preview)
PowerShell example:
$env:AZURE_OPENAI_MODEL="your-deployment-name"
$env:AZURE_OPENAI_API_KEY="your-api-key"
$env:AZURE_OPENAI_ENDPOINT="https://your-resource-name.openai.azure.com/"
$env:AZURE_OPENAI_API_VERSION="2024-07-01-preview"
If you use a non-Azure model API, keep this contract unchanged:
ActionGen.query(...)must return one text command each step.- Returned command should still be compatible with
parse_llm_action(...).
Minimal expected return format:
Thinking: <your model reasoning>
Command: move_forth
Experimental Results
Quantitative Results
We evaluate 17 representative models across five categories: Basic Baselines, Non-Reasoning LMMs, Reasoning LMMs, Agent-Based Approaches, and Vision-Language-Action Models.
Note: Short, Middle, and Long groups correspond to ground truth trajectories of <118.2m, 118.2-223.6m, and >223.6m respectively. SR = Success Rate, SPL = Success weighted by Path Length, DTG = Distance to Goal.