ac4462/SceneDescriptionNuPlan

Scene Description NuPlan Dataset

A dataset of natural language descriptions generated from autonomous driving simulator states, paired with visualization images of driving scenarios.

Overview

The scene_description_nuplan_dataset contains natural language descriptions of driving scenarios generated from the V-Max simulator using the NuPlan dataset. This dataset provides paired data of:

1. Natural language descriptions of driving scenes
2. Corresponding visualization images at different timesteps
3. Multiple types of representations of the same scene

Rule-Based Approach

We chose a rule-based approach to generate driving scene descriptions because it offers precise control over content and format while ensuring factual accuracy by directly translating simulator states into language. This method eliminates the need for expensive human annotation while allowing for flexible focus on different driving aspects. The approach provides systematic coverage of all relevant driving elements and ensures reproducibility through its deterministic nature. Ultimately, this creates a reliable dataset bridging visual driving scenes and natural language, valuable for training multimodal AI systems for autonomous driving.

Dataset Structure

scene_description_nuplan_dataset/
├── scene_descriptions.json       # JSON file containing all scenario descriptions
├── images/                       # Main visualization images of driving scenarios
│   ├── scenario_0_step_0.png
│   ├── scenario_0_step_1.png
│   └── ...
├── simulator_state_images/       # Images visualizing the simulator state
│   ├── simulator_state_0_step_0.png
│   └── ...
├── observation_images/           # Images showing agent observations
│   ├── observation_0_step_0.png
│   └── ...
└── features_images/              # Images visualizing extracted features
    ├── features_0_step_0.png
    └── ...

Description Format

Each scenario in the dataset contains:

• scenario_id: Unique identifier for the scenario
• descriptions: Array of natural language descriptions for each timestep
• image_files: Array of filenames for the main visualization images
• num_steps: Total number of timesteps in the scenario

Example Description

Natural language descriptions include information about:

• Ego vehicle state (speed, position)
• Nearby agents (vehicles, pedestrians, cyclists)
• Traffic lights and their states
• Road features
• Planned route information
• Upcoming turns or maneuvers

Example description:

The ego vehicle is moving at 7.5 m/s.
There are 1 vehicle(s) nearby.
The closest vehicle is 9.2m in front and 2.9m to the right and is moving at 8.4 m/s.
The ego vehicle is following its planned route.
The ego vehicle is on a straight section of its route.

Dataset Generation

This dataset was generated using:

1. The V-Max simulator framework
2. NuPlan dataset scenarios as input
3. A specialized natural language generation system that converts simulator states into descriptive text
4. Expert policy for agent behavior

The system analyzes various aspects of the driving scene including:

• Vehicle positions, speeds, and headings
• Roadgraph elements (lanes, boundaries)
• Traffic lights and signals
• Planned routes and trajectories

Usage

This dataset can be used for:

1. Training vision-language models for autonomous driving
2. Developing natural language interfaces for autonomous vehicles
3. Research on explainable AI for autonomous driving
4. Scene understanding and situation awareness in driving contexts
5. Multimodal learning combining visual and textual representations of driving scenarios

Dataset Statistics

• Number of scenarios: 50
• Total number of timesteps/descriptions: ~800
• Average number of timesteps per scenario: ~16
• Types of scenarios: Urban driving situations including intersections, traffic lights, interactions with other road users

Image Types

The dataset includes multiple visualizations for each scenario step:

1. Main images: Top-down view of the driving scene with agents and roadmap (simulator state, Observation, and features combined)
2. Simulator state images: Visualization of the v-max simulator state
3. Observation images: Representation of the ego vehicle's observations
4. Feature images: Visualization of extracted features can be used for decision making