PaperlayerView exact revision on arXiv

Source-linked AI summary

AuRaSim: A Ray-Tracing Based Automotive FMCW Radar Simulator for Dynamic Scenarios

Zhaoze Wang, Changxu Zhang, Tai Fei, Christopher Grimm, Yi Jin, Claas Tebruegge, Markus Gardill

arXiv:2610.37403v1cs.ROeess.SP

TL;DR

Automotive radar simulation remains challenging because realistic measurement synthesis combines dynamic scenes, electromagnetic propagation, and coherent signal formation, while existing options can be proprietary or simplified. AuRaSim builds on Blender and Sionna RT to simulate dynamic scenarios and generate coherent multi-channel ADC signals. In digital-twin evaluations, simulated and measured range–Doppler maps achieved average RMSE 0.147 and PSNR 16.93 dB.

  • Problem

    Realistic automotive radar simulation must combine dynamic 3D scenes, electromagnetic materials and propagation, and multi-channel signal formation, while existing tools may be simplified or proprietary.

    Source: §I · p. 1

  • Method

    AuRaSim extends Blender and Sionna RT with radar-specific propagation and signal synthesis for dynamic scenarios.

    Source: §I · p. 2

  • Results

    Average RMSE 0.147 and PSNR 16.93 dB were obtained when comparing simulated and measured range–Doppler maps in digital-twin scenarios.

    Source: §VI · p. 14

  • Takeaways & Limitations

    AuRaSim provides a common, traceable scene-to-signal pipeline that retains a physically interpretable mapping from scene to radar data.

    Source: §VI · p. 14

Abstract

from arXiv · show

Modern simulators can generate synthetic camera and LiDAR data with accurate annotations for downstream tasks. In contrast, automotive radar simulation remains less mature because it requires not only a 3D engine to model dynamic objects in the scenes but also electromagnetic material properties, antenna configurations, wave propagation mechanisms, and coherent signal synthesis. Existing frameworks are either proprietary or overly simplified, limiting the scalability of radar raw measurements. This paper presents AuRaSim, a ray-tracing-based automotive frequency-modulated continuous wave (FMCW) radar simulation framework for dynamic scenarios. Rather than implementing all components from scratch, AuRaSim builds on Blender for efficient dynamic-scene construction and Sionna RT for graphics processing unit (GPU) accelerated electromagnetic ray tracing, while introducing radar-specific adaptations for path-resolved propagation with complex coefficients, motion-aware Doppler modeling, and subsequently coherent multi-channel analog-to-digital converter (ADC) cube synthesis based on waveform configurations. Quantitative comparison between simulated and measured range-Doppler maps in digital-twin scenarios achieves a root-mean-square error (RMSE) of 0.147 and a peak signal-to-noise ratio (PSNR) of 16.93 dB, demonstrating the accuracy and fidelity of the synthetic signal. AuRaSim therefore provides a complete and extensible platform for reproducible automotive radar simulation and research. Project homepage: https://wangzhaoze.github.io/aurasim.

I. Introduction

Automotive radar simulation must balance realistic dynamic scenes, electromagnetic propagation, and signal synthesis, yet existing approaches are limited by scale, abstraction, or access. AuRaSim combines open-source scene construction and ray tracing into an extensible scenario-to-signal framework for dynamic automotive radar.

I. Introduction

  • AuRaSim integrates Blender and Sionna RT into an end-to-end automotive FMCW radar simulation framework, extending both rather than building the scene and propagation stack from scratch.Blender supports complex polygonal scenes and animation; Sionna RT provides electromagnetic material models and ray tracing.Source: §I · p. 2

A. Physics-Based Radar Simulation

  • Physics-based automotive radar simulators model measurements from scene, hardware, and propagation, but prior work varies in dynamic-scene coverage, Doppler treatment, and output level.AuRaSim’s stated distinction is a unified pipeline for dynamic scenes, physics-based propagation, and coherent multi-channel ADC synthesis.Source: A. Physics-Based Radar Simulation · p. 2

B. Generative Radar Synthesis

  • Data-driven radar synthesis spans representations from radar images to raw chirps, but learned distributions can limit generalization to customized antenna arrays and waveform parameters.AuRaSim contrasts this with configurable hardware and physics-based coherent multi-channel ADC generation.Source: B. Generative Radar Synthesis · p. 3

C. Radar Neural Fields

  • Radar neural-field methods render novel views of reconstructed scenes, whereas AuRaSim simulates newly constructed dynamic driving scenarios with path-resolved propagation.The cited approaches include methods restricted to static scenes or 2-D range–azimuth inputs and outputs.Source: C. Radar Neural Fields · p. 3

D. Automotive Radar Simulation Software

  • Radar simulation software ranges from detection-level models to physically based tools, but proprietary or licensed implementations can restrict inspection and modification.AuRaSim is presented as an accessible and extensible research framework for automotive FMCW radar studies.Source: D. Automotive Radar Simulation Software · p. 3

III. Preliminaries

The preliminaries explain how FMCW signals encode range, Doppler, and angular information, then outline the path solver and ray initialization used in AuRaSim. Spherical Fibonacci sampling provides approximately equal-area coverage of the sphere.

A. FMCW Radar Signal Processing

  • Fast-time phase evolution carries range information, slow-time phase progression carries Doppler, and relative phase across antenna channels carries angular information in the coherent ADC cube.The cube coherently sums beat-signal contributions from resolved paths while preserving relative phase across samples, chirps, and channels.Source: A. FMCW Radar Signal Processing · pp. 3–4

B. Sionna RT Path Solver

  • AuRaSim uses Sionna RT’s propagation formulation as the basis for its path solver, with further implementation details referenced to the Sionna RT work.The section also identifies specular reflection, diffuse scattering, thin-slab transmission, and edge diffraction as principal propagation mechanisms.Source: B. Sionna RT Path Solver · p. 4, Figure 1 · p. 4

1) Ray Sampling:

  • Spherical Fibonacci sampling initializes rays with an approximately equal-area distribution over the sphere, reducing polar clustering compared with constant angular increments.The angles are defined for Nray rays, and each initial ray tube represents an approximately equal solid angle.Source: 1) Ray Sampling · p. 4

2) SBR Candidate Generation:

The SBR stage samples candidate interaction sequences to avoid branching at every intersection, then models how each interaction changes propagation geometry and the complex electromagnetic field.

2) SBR Candidate Generation:

  • At each propagation depth, SBR samples at most one interaction type per ray, limiting the exponential growth from spawning every possible continuation.Specular reflection, diffuse scattering, and transmission are importance-sampled from material response and Fresnel coefficients; diffraction receives a separate sampling probability to explore weaker paths.Source: 2) SBR Candidate Generation · p. 4

3) Electromagnetic Propagation Mechanisms:

  • Specular reflection applies a polarization-dependent complex Fresnel response, while diffuse scattering distributes reflected energy according to a scattering pattern.The Lambertian pattern is the default; directive and backscattering patterns can concentrate energy toward the specular direction or incident direction, respectively.Source: 3) Electromagnetic Propagation… · p. 5
  • Diffuse interactions use next-event estimation to test receiver visibility, while stochastic continuation and ray-tube weights normalize sampled path contributions.This retains a receiver-directed candidate when the receiver is visible and continues sampling to explore higher-order paths.Source: 3) Electromagnetic Propagation… · p. 5
  • Transmission uses a homogeneous thin-slab model with internal reflections, and edge diffraction uses UTD with only first-order diffraction considered.The slab coefficient accounts for both interfaces and propagation through the slab; the diffraction coefficient determines complex field amplitude and phase.Source: 3) Electromagnetic Propagation… · p. 5

4) Image-Method Path Refinement:

Image-method refinement turns sampled interaction sequences into geometrically valid transmitter-to-receiver paths. The solver then evaluates complex fields along those paths using antenna responses, interaction matrices, and path-dependent spreading.

4) Image-Method Path Refinement:

  • SBR generates candidate interaction sequences, and image-method refinement finds exact interaction points that connect the transmitter and receiver while satisfying propagation constraints.Candidates with interaction points outside their primitives or occluded paths are discarded.Source: 4) Image-Method Path Refinement · p. 5

5) Electromagnetic Field Calculation:

  • For each validated path, the solver applies complex polarization-dependent interaction matrices in propagation order, evaluates the receive antenna response, and accounts for geometric spreading.Path-dependent spreading includes stochastic-sampling normalization, additional interaction and ray-tube weights for diffuse paths, and UTD spreading for diffracted paths.Source: 5) Electromagnetic Field Calculation · pp. 5–6

IV. Radar Simulation System

AuRaSim links configurable radar sensors and dynamic Blender scenes to a path-solving stage and coherent multi-channel ADC synthesis. Its modular stages preserve traceability and can be inspected and reused independently.

A. System Overview

  • AuRaSim maps a dynamic digital-twin scene and configurable radar sensor to multi-channel ADC signals through path resolution and waveform-dependent synthesis.The path set retains complex coefficients, delays, Doppler frequencies, and propagation geometry; retaining it decouples propagation from waveform-dependent synthesis.Source: A. System Overview · p. 6

B. Radar Sensor Configuration

  • The radar sensor configuration combines MIMO layout, antenna radiation patterns, and waveform parameters.The layout supports explicit-array tracing or reference-array approximation; TDM-MIMO retains transmit sequence to preserve virtual-array channel timing.Source: B. Radar Sensor Configuration · p. 6

2) Antenna Pattern:

  • Customized antenna patterns are interpolated at ray-tracing directions and incorporated into electromagnetic field calculations.The pattern is represented as a direction-dependent complex field response F(θ, ϕ).Source: 2) Antenna Pattern · p. 6

3) Waveform Parameters:

  • FMCW waveform configuration includes carrier frequency, bandwidth, chirp slope and durations, ADC sampling frequency, samples per chirp, and chirps per frame.AuRaSim can synthesize signals for different waveforms by specifying the transmit waveform and sampling scheme.Source: 3) Waveform Parameters · p. 6

C. Blender Scenario Builder

  • The Blender Scenario Builder constructs plausible dynamic driving scenes from trajectory-level descriptions.It supplies the downstream Path Solver with synchronized geometry and mesh motion state.Source: C. Blender Scenario Builder · p. 6

1) MIMO Configuration:

The scene builder represents static, rigid, and deformable geometry and updates these categories persistently. A world-space vertex-velocity field supplies local interaction velocities for motion-aware Doppler modeling.

1) Dynamic Scene Representation and Persistent Update:

  • Static infrastructure, rigid dynamic objects, and articulated or deformable objects are represented as separate scene-geometry categories.Examples include roads and buildings, vehicles, and pedestrians, respectively.Source: 1) Dynamic Scene Representation and… · p. 7
  • Persistent scene updates instantiate static geometry once, update rigid objects through transformations, and update deformable objects' vertex positions while retaining triangle connectivity.This avoids rebuilding and transferring the complete electromagnetic scene to the GPU for every radar frame.Source: 1) Dynamic Scene Representation and… · p. 7

2) Radio Material Assignment:

  • Radio materials assigned to meshes specify electromagnetic properties used in material-dependent propagation interactions.The properties include permittivity, conductivity, thickness, diffuse scattering, cross-polarization, and scattering pattern; built-in ITU models are evaluated at the configured radar frequency.Source: 2) Radio Material Assignment · p. 7

3) Vertex-Velocity Field:

  • A world-space vertex-velocity field captures local motion on rotating, articulated, or deformable objects for hit-point Doppler modeling.Rigid-object vertex velocities combine translation and rotation, while evaluated articulated or deformable meshes use finite differences; ray-interaction velocities are interpolated from triangle-vertex velocities.Source: 3) Vertex-Velocity Field · p. 7

D. Path Solver

AuRaSim adapts Sionna RT's Path Solver with radar-oriented ray sampling and explicit path handling. FOV-aware Fibonacci sampling concentrates a fixed ray budget in the radar-relevant region while retaining the represented electromagnetic energy.

D. Path Solver

  • The Path Solver combines shooting-and-bouncing-rays with image-method refinement, then evaluates complex electromagnetic fields and exports explicit paths for signal synthesis.Its radar-specific adaptations include ray sampling and filtering, hit-point motion information, and path-level representation.Source: D. Path Solver · p. 8, Figure 4 · p. 8

1) FOV Ray Sampling:

  • For the same number of rays, FOV-aware Fibonacci sampling increases angular sampling density within the radar field of view.Matching that in-FOV density with full-sphere sampling would require substantially more rays and higher ray-tracing cost.Source: 1) FOV Ray Sampling · p. 8, Figure 4 · p. 8
  • FOV-restricted sampling updates the per-ray solid-angle weight, improving sampling efficiency without artificially changing represented electromagnetic energy.The figure contrasts nonuniform point density from uniform angular sampling, more uniform full-sphere spherical Fibonacci sampling, and concentrated FOV-aware sampling.Source: 1) FOV Ray Sampling · p. 8, Figure 4 · p. 8

2) Path Filtering:

The Path Solver filters paths against the radar's receive field of view and maximum distance before refinement and field evaluation. Hit-point Doppler models local interaction motion, and explicit path records can be reused for different waveform and acquisition configurations.

2) Path Filtering:

  • Paths are retained only when their arrival directions fall within the receive field of view and their propagation distances do not exceed the configured maximum.Discarding paths outside either condition avoids unnecessary path refinement and field evaluation.Source: 2) Path Filtering · p. 8

3) Hit-Point Doppler:

  • AuRaSim replaces object-level velocity with local hit-point velocity when calculating first-order path Doppler.Each moving interaction point contributes according to its velocity projected onto the change between outgoing and incoming propagation directions, allowing distinct Doppler shifts at different locations on one mesh.Source: 3) Hit-Point Doppler · p. 9

4) Path Representation:

  • Explicit path-level records separate propagation from waveform-dependent signal generation and can be reused across waveform and acquisition configurations.The records provide input to the ADC Signal Synthesizer.Source: 4) Path Representation · p. 9

E. ADC Signal Synthesizer

AuRaSim converts resolved propagation paths into coherent multi-channel FMCW ADC signals. It supports explicit or synthetic array modeling and represents motion within frames through Doppler phase progression while updating scene velocities between frames.

E. ADC Signal Synthesizer

  • Coherent superposition of resolved paths preserves phase and captures constructive and destructive interference between multipath components.The synthesizer forms complex sampled FMCW beat signals rather than adding received powers independently.Source: E. ADC Signal Synthesizer · p. 9

1) MIMO Array Modeling:

  • Synthetic-array mode reconstructs element-dependent phase differences from path departure and arrival directions, while explicit-array mode traces each transmit–receive pair.The synthetic formulation assumes a planar wavefront across the aperture; explicit tracing can be used for significant near-field effects at increased computational cost.Source: 1) MIMO Array Modeling · p. 9

2) Coherent Multipath Synthesis:

  • Within each frame, Doppler-induced phase progression represents motion across chirps; velocities update between frames to simulate time-varying articulated-object micro-Doppler.Frame-level path evaluation reduces computational cost; chirp-level path evaluation is also supported at increased ray-tracing time.Source: 2) Coherent Multipath Synthesis · p. 9

A. Data Preparation

The evaluation uses synchronized real radar and RGB data to build digital-twin scenarios, then compares simulated and measured range–Doppler maps under matched radar configurations. RMSE and PSNR quantify map agreement.

1) Dataset:

  • The dataset contains approximately 19,800 synchronized raw radar frames and RGB images from campus, road, and parking-lot settings, covering pedestrians, cyclists, and cars.The radar configuration used for the real measurements and AuRaSim is summarized in Table I.Source: 1) Dataset · p. 9, Table I · p. 10

2) Digital-Twin Reconstruction:

  • Digital twins combine manually created, radar-aligned static environments with detected, tracked, and smoothed dynamic-object trajectories animated using 3D assets.The reconstructed scenes are used for real-to-sim validation and ablation studies.Source: 2) Digital-Twin Reconstruction · p. 10

3) Parameters:

  • The measurements use a 77-GHz Texas Instruments AWR1843; AuRaSim and RadarSimPy reproduce the same reconstructed scene, antenna, waveform, and sampling configuration.RadarSimPy's ray density is set to 0.1 rays per wavelength squared, the highest stable tested setting for these scenes; the main AuRaSim experiments launch 10 million rays within ±90° azimuth and elevation, with maximum interaction depth 3.Source: 3) Parameters · p. 10

B. Evaluation Metrics

  • RMSE and PSNR compare normalized log-power values across range–Doppler bins, with lower RMSE and higher PSNR indicating closer agreement with measured radar response.Measured and simulated maps use the same processing pipeline, are independently normalized, clipped from −60 dB to 0 dB, and mapped to [0, 1].Source: B. Evaluation Metrics · p. 10

C. Qualitative Results

AuRaSim's simulated range–Doppler maps more closely match measured structures than the baseline in the reported comparisons, and its dynamic-scene simulation captures articulated-motion signatures. Ablations show a ray-count accuracy–runtime trade-off in the tested scene.

C. Qualitative Results

  • AuRaSim's target responses more closely resemble measured range–Doppler structures, especially the extent and magnitude distributions of extended targets and scenes with multiple moving objects.The difference is most apparent in MLMS001, which contains more dynamic objects; Fig. 5 compares real and rendered views with RadarSimPy, AuRaSim, and measured maps.Source: C. Qualitative Results · p. 10, Figure 5 · p. 11
  • Periodic limb motion produces time-varying Doppler spreads around the pedestrian's main body return, with larger limb-velocity excursions generating outer spectral components.The torso remains concentrated near the dominant body velocity, showing local body-part motion rather than a single rigid-target representation.Source: C. Qualitative Results · p. 10, Figure 6 · p. 12

D. Quantitative Results

  • Across three sequences, AuRaSim reduces RMSE by approximately 38.8% and improves PSNR by 3.50 dB versus RadarSimPy; MLMS001 shows RMSE 0.233 to 0.111 and PSNR 12.90 to 19.40 dB.The comparison is end-to-end under matched reconstructed scenes and radar configurations, not an isolated comparison of ray-tracing accuracy.Source: D. Quantitative Results · pp. 10–11, Table II · p. 12

E. Ablation Study

  • The ablation comparisons apply to their respective experimental configurations rather than serving as absolute simulator benchmarks.The study examines ray count, FOV-aware sampling, path filtering, local hit-point Doppler, and MIMO array modeling; computational cost and reconstruction quality also depend on settings and reconstruction quality.Source: E. Ablation Study · pp. 11–12

1) Effect of the Number of Rays:

  • Increasing launched rays from 0.1 to 10 million raises PSNR from 16.81 dB to 19.40 dB and runtime from 0.672 s/frame to 0.838 s/frame in MLMS001.Beyond sufficiently sampling the scene's dominant geometric interactions, gains diminish while ray-tracing cost continues to rise; 5–10 million rays gives a dense representation for this scene.Source: 1) Effect of the Number of Rays · p. 12, Table III · p. 13

2) Effect of FOV Ray Sampling:

AuRaSim’s ablations show that radar-focused ray sampling, path filtering, and hit-point Doppler modeling improve simulated radar responses, while synthetic-array modeling reduces ray-tracing time with comparable range–azimuth fidelity. Across evaluated sequences, the simulator’s average RMSE is 0.147 and its average result is 16.93 dB.

2) Effect of FOV Ray Sampling:

  • 18.71 dB versus 18.59 dB with 5 million versus 10 million rays, while runtime falls from 0.859 to 0.774 s/frame (9.9%) under FOV-restricted sampling.At an identical 10-million-ray budget, FOV-aware sampling raises PSNR from 18.59 to 19.40 dB and reduces runtime from 0.859 to 0.838 s/frame by avoiding rays outside the radar FOV.Source: 2) Effect of FOV Ray Sampling · p. 13

3) Effect of Path Filtering:

  • Removing path filtering drops PSNR from 19.40 to 15.46 dB, while runtime decreases from 0.838 to 0.826 s/frame.Responses from paths outside the radar observation region, particularly around zero Doppler, reduce relative contrast without filtering.Source: 3) Effect of Path Filtering · p. 13

4) Effect of Doppler Modeling:

  • 17.52 to 19.40 dB with runtime essentially unchanged (0.837 versus 0.838 s/frame) when using hit-point rather than object-level Doppler.Spatially varying hit-point velocities produce a more realistic wheel-induced micro-Doppler distribution that better agrees with measured radar response.Source: 4) Effect of Doppler Modeling · p. 13

5) Effect of MIMO Array Modeling:

  • Below 0.2 dB PSNR difference and approximately 58% less ray-tracing time distinguish synthetic-array from explicit-array modeling in the evaluated setting.The comparison uses compact arrays and target distances where the far-field approximation applies; larger apertures or nearby targets may introduce significant wavefront curvature and channel-dependent delays.Source: Table IV · p. 14, Figure 8 · p. 14, 5) Effect of MIMO Array Modeling · p. 14

VI. Conclusion

  • AuRaSim connects dynamic digital-twin scenes to coherent waveform-level radar synthesis in a traceable pipeline, with material accuracy and dense-scene computation remaining constraints.Future work targets measurement-driven calibration, more complete environmental modeling, faster ray tracing and ADC synthesis, and applications including differentiable inverse radar rendering and synthetic-data generation.Source: §VI · p. 14
  • 0.147 average RMSE and 16.93 dB average across evaluated sequences quantify AuRaSim’s comparison with measured range–Doppler maps.The authors report that the simulated maps reproduce dominant measured response structures while retaining a physically interpretable scene-to-signal mapping.Source: §VI · p. 14

Related papers

Loading 2610.37403v1…