Skip to content

Simulator — DIVP

RealSim-CP is generated with DIVP (Driving Intelligence Validation Platform), distributed as V-Drive for MATLAB/Simulink®. DIVP is a physics-based sensor-simulation system that uses ray tracing and electromagnetic-wave modeling to produce high-fidelity camera, LiDAR, and radar data comparable to real-world quality.

Official site

DIVP is developed under Japan's SIP (Strategic Innovation Promotion) program. Product information: https://divp.net/. This page summarizes how RealSim-CP uses it; it is not a substitute for the official V-Drive documentation.

Why simulation

Real-world cooperative-perception data is expensive and slow to collect — it requires many instrumented agents and large-scale multimodal labeling. DIVP produces large volumes of synchronized multimodal data across diverse scenarios and weather/time conditions at far lower cost, while staying close to real sensor physics through calibration against real measurements.

How it compares

Many simulators generate LiDAR by approximating object shapes and positions. DIVP instead physically simulates camera and LiDAR operation from electromagnetic-wave principles, renders spatially with ray tracing, and improves accuracy through calibration with real-world experimental data.

Sensor models

  • Camera — models the spectral response incident on the CMOS sensor (rather than emitting RGB directly), with solar illumination from sky models, Monte Carlo ray-traced spatial rendering, and lens distortion. RealSim-CP uses a Sony IMX490 model with a 120° horizontal field of view.
  • LiDAR — ray traces signal propagation and solves the physical equations for received intensity, reproducing effects such as backlighting. RealSim-CP uses a Velodyne VLS-128 model.

In Simulink, DIVP is organized as blocks, one per sensing/processing stage — for example the camera chain (CameraRendering, CameraPerception, ToRGB, PostFog, …), the LiDAR chain (LidarRendering, transmitter/receiver pairs, LidarV10/LidarV20), radar, and scenario blocks (Scenario, ScenarioReader, ScenarioMerge, TrafficLight).

Generation pipeline

flowchart TB
  subgraph design["1 - Scenario design (SDMG)"]
    sdmg["Space Design Model Generator"] --> files["scenario XML + trajectory CSVs<br/>maps, agents, sensors, behaviors"]
  end
  subgraph run["2 - Simulation (V-Drive)"]
    files --> vdrive["V-Drive in MATLAB/Simulink"]
    vdrive --> out["synchronized images + point clouds @ 10 Hz"]
  end
  subgraph multi["3 - Cooperative assembly"]
    out --> decomp["per-agent single scenarios<br/>simulated independently"]
    decomp --> combine["combine + annotate"]
    combine --> ol["OpenLABEL dataset"]
  end
  1. Scenario design (SDMG). The Space Design Model Generator produces the scenario definition (XML) and trajectory data (CSV): maps, agent configurations, sensor parameters, and behaviors.
  2. Simulation (V-Drive). The scenario is imported into V-Drive, which builds the environment, runs the camera and LiDAR physics, and outputs synchronized image and point-cloud data at 10 Hz.
  3. Cooperative assembly. DIVP does not natively support V2X scenarios, so a multi-agent scene is decomposed into several single-agent scenarios. Each sensor-equipped agent is simulated independently, then the outputs are recombined and annotated in OpenLABEL. A Python automation tool manages and sequentially runs the scenarios via the MATLAB engine.

Performance

Generation used a virtual machine with an NVIDIA RTX 6000 Ada GPU. Rendering a single 1158 × 750 image takes roughly 5–10 seconds in clear weather and 5–10 minutes under rain. Camera and LiDAR data are recorded synchronously at 10 frames per second.

Acknowledgement

The dataset was generated with the DIVP® simulator. We thank the DIVP consortium for providing access to the platform. This work was supported by JST CRONOS, Japan (Grant Number JPMJCS24K8).