SimART: 6G ISAC Simulation Platform

• SimART is a unified simulation platform for 6G ISAC that integrates robotics, ray tracing, and wireless modules via synchronized ROS-based streams.
• It decouples sensing and physics front ends from the wireless back end using a shared simulation clock and unified coordinate frame, ensuring reproducibility.
• The platform supports versatile scene construction and evaluates high-accuracy beam prediction, demonstrating robust performance under diverse conditions.

SimART is a unified, open simulation platform designed explicitly for sixth-generation integrated sensing and communication (6G ISAC). It integrates mature robotics, ray tracing, and wireless evaluation engines into a single reproducible pipeline centered on a robot operating system (ROS) backbone, with the stated objective of producing multimodal datasets and experiments that jointly characterize platform mobility, onboard sensing, site-specific wireless propagation, and link/system-level performance (Yan et al., 13 May 2026).

1. Definition and research context

SimART addresses a gap in multimodal ISAC research identified in the paper: existing tools typically model only subsets of the required stack. Standalone robotics simulators such as CARLA, AirSim, Gazebo, and Isaac Sim provide physics, sensors, and ROS integration, but they do not provide site-specific wireless models, link/system-level evaluation, or channel knowledge map (CKM) construction. Ray-tracing and wireless data generators such as Sionna RT, DeepMIMO, and Wireless InSite can generate channels or limited performance metrics, but they generally lack realistic platform motion, onboard sensors, and ROS-based integration. Network simulators such as ns-3 operate at a higher abstraction level and do not represent 3D geometry, platform dynamics, or multimodal sensing streams (Yan et al., 13 May 2026).

Against that background, SimART is presented not as another standalone simulator but as a ROS-centered integration layer. Its key novelty lies in synchronizing all multimodal streams through a shared simulation clock, a unified coordinate frame implemented as a tf2 tree, and timestamped messages, while decoupling sensing and physics front ends from the wireless back end. This decoupling allows ROS-compatible simulators to be exchanged while reusing the same ray-tracing, link/system, and CKM pipeline across aerial, ground, indoor, and maritime ISAC scenarios. The platform also adds a dual scene construction pipeline, covering both real-world OpenStreetMap-based scenes and user-defined layouts, and a CKM generator that aggregates ray-tracing and link/system outputs into location-specific priors for ISAC algorithms (Yan et al., 13 May 2026).

2. ROS-centered architecture and multimodal synchronization

The architecture is organized into four functional modules exposed through unified ROS interfaces: Physics and Sensing, Ray Tracing, Link and System, and CKM generator. These modules run as concurrent ROS nodes and exchange standardized messages. The Physics and Sensing module is defined by a ROS contract and publishes platform pose on the tf2 tree together with standard ROS sensor topics including RGB, depth, semantic segmentation, LiDAR point clouds, IMU, GNSS/GPS, and ground-truth pose. AirSim is used as the reference front end, while Gazebo, Isaac Sim, and CARLA are described as ROS-compatible alternatives under the same contract (Yan et al., 13 May 2026).

The synchronization mechanism is central to SimART. All modules run under ROS simulated time. A dedicated clock node advances the simulation time independently of wall-clock speed, so alignment is preserved even when ray tracing runs slower than real time. Spatial consistency is enforced through a common coordinate frame in which world, transmitter, receiver, and sensor frames are registered in a tf2 frame tree rooted at a configurable world origin. Every message carries a header timestamp from the shared clock, and ROS approximate time synchronization is used to keep streams aligned (Yan et al., 13 May 2026).

A notable consequence of this design is the unification of recording and replay. A single rosbag captures the full session, including sensors, channels, link-level KPIs, CKM layers, and frame transforms. Standard ROS tools such as rqt_bag and rviz can then be used for synchronized inspection. The paper emphasizes that this removes the need for manual timestamp and coordinate alignment and makes experiments reproducible at the level of a single file (Yan et al., 13 May 2026).

3. Scene construction, propagation assets, and wireless evaluation

SimART produces two spatially aligned assets for each scene: a high-fidelity visual asset for the physics and sensing module and a simplified propagation-ready mesh for ray tracing. Both assets share a world origin so that transceiver and sensor poses remain consistent across modules. In the real-world pipeline, the input is an OpenStreetMap extract containing building footprints with heights, roads, and land-use. The visual branch uses OSM2World to extrude footprints and generate roads and vegetation, which are then imported into Unreal Engine as a static level used by AirSim or another ROS-compatible simulator. The wireless branch loads the same OpenStreetMap data into Blender, assigns electromagnetic material properties such as relative permittivity and conductivity, and exports a Mitsuba scene for Sionna RT (Yan et al., 13 May 2026).

The user-defined scene interface supports visual scenes built in RoadRunner or Unreal Editor. SimART’s Blender script then converts such scenes into a propagation-ready mesh by removing fine geometric details such as window frames and railings based on tags and size thresholds, applying quadric edge-collapse decimation while preserving sharp edges along facades and ground planes, and assigning materials from Sionna RT’s library. The result is reported as one-to-two orders of magnitude fewer triangles, while dominant specular and diffraction paths are preserved. The paper explicitly characterizes this as a trade-off: mesh simplification removes details that contribute mainly diffuse scattering, producing significant ray-tracing speedups at the cost of some channel-realism loss (Yan et al., 13 May 2026).

The wireless back end consists of two linked engines. The Ray Tracing module is built on Sionna RT and computes per-path quantities including complex amplitude, delay, angle of arrival, Doppler, and interaction points, which are then aggregated into channel impulse responses (CIRs). It supports three modes: online, driven by live poses; offline replay, driven by a trajectory rosbag; and grid scan, which evaluates a dense virtual receiver grid for CKM generation. The Link and System module is built on Sionna SYS and ingests CIRs to simulate PHY/MAC layers with configurable OFDM numerology, multi-antenna chains, and digital beamforming codebooks. It outputs KPIs such as SINR, BLER, and achievable rate, and for beam management it reports the optimal codebook beam maximizing SINR at each time instant (Yan et al., 13 May 2026).

For a narrowband snapshot, the paper expresses the site-specific MIMO-OFDM channel model in the form

$y = \sqrt{P}\, H\, w\, s + n,$

where $H$ is the MIMO channel, $w$ is the transmit beamforming vector from a codebook, $s$ is the transmit symbol, $P$ is the transmit power, and $n$ is noise. The optimal beam is defined as

$b^* = \arg\max_b \mathrm{SINR}_b.$

These quantities are subsequently reused by the CKM machinery and by downstream ISAC tasks (Yan et al., 13 May 2026).

4. Channel Knowledge Maps

A CKM in SimART is defined as a set of spatial priors over a region of interest, computed deterministically through ray tracing and link/system evaluation at each grid cell. The implementation discretizes the region of interest as a 2D ground-plane grid at a configurable receiver height. With the transmitter fixed, virtual receivers are placed at grid centers, and each cell is evaluated by ray tracing and system-level simulation. The resulting map is described by

$$K(r) = \{ PL(r), \tau_{rms}(r), \sigma_\theta(r), \sigma_\phi(r), SINR(r), R(r), b^*(r) \},$$

where $r \in \mathbb{R}^2$ denotes ground-plane coordinates at fixed receiver height $z_0$ (Yan et al., 13 May 2026).

The channel impulse response at location $H$0 is given as

$H$1

with $H$2 indexing paths, $H$3 the complex amplitudes, and $H$4 the delays. From this representation, the paper defines path loss as

$H$5

relative to transmit power normalization, and RMS delay spread as

$H$6

Angular spreads are derived from per-path angle-of-arrival and angle-of-departure quantities, while the Link and System module computes $H$7, achievable rate $H$8, and the optimal beam index $H$9 under configured numerology, antennas, and codebooks (Yan et al., 13 May 2026).

The paper stresses that CKMs are deterministic with respect to the ray-tracing and link-level models and that their generation uses exhaustive grid scanning with configurable resolution. No interpolation is required for CKM construction itself, although downstream algorithms may interpolate afterward if desired. Example layers include path loss, received power, best-BS achievable rate, and effective SINR. Within the platform, CKMs function as spatial priors for ISAC algorithms and as tools for dataset analysis and failure diagnosis (Yan et al., 13 May 2026).

A stated limitation is that the current CKM is 2D, with a fixed receiver height. Extension to 3D by multi-height sweeps is identified in the paper as a natural future direction (Yan et al., 13 May 2026).

5. Vision- and position-aided beam prediction case study

The paper’s case study uses SimART to evaluate vision- and position-aided beam prediction. The scenario consists of a rooftop base station at 27 m AGL equipped with an $w$0 uniform planar array operating at 3.5 GHz, serving a single-stream UAV through a 64-beam DFT codebook. UAV altitude varies between 1 m and 20 m along predefined trajectories through line-of-sight areas relative to the base station. An RGB camera collocated with the base station observes the airspace, and a YOLOv8 detector identifies the UAV in each frame, producing image-domain location; this is combined with synchronized GNSS/GPS position to predict the optimal transmit beam index (Yan et al., 13 May 2026).

The scene is user-defined in RoadRunner and converted into paired AirSim and Sionna RT assets with a shared coordinate frame. Three environmental conditions are used: sunny daytime, rainy daytime, and nighttime. Geometry and trajectories are kept identical, while AirSim changes illumination and weather and Sionna RT modifies materials for dry versus wet surfaces. The labels are the optimal beam indices from the 64-beam DFT codebook, defined as the beam achieving $w$1 at each timestamp. Evaluation uses Top-1, Top-3, and Top-5 accuracy on beam index prediction; latency is described as being governed by ROS pipelines and model inference, with synchronized timestamps enabling consistent comparisons (Yan et al., 13 May 2026).

The reported quantitative results are as follows. Under sunny daytime conditions, the method attains Top-1 97.86%, Top-3 99.83%, and Top-5 100.00%. Under rainy daytime conditions, it attains Top-1 95.39%, Top-3 99.84%, and Top-5 100.00%. Under nighttime conditions, it attains Top-1 97.92%, Top-3 99.94%, and Top-5 100.00% (Yan et al., 13 May 2026).

The paper’s discussion attributes the high accuracies to strong spatial cues from vision detections and GPS, together with site-specific beam labels derived from ray-traced channels. It also notes a moderate Top-1 degradation in rain, associated with visual appearance changes and slightly altered propagation and material responses, while Top-3 and Top-5 remain near-perfect. CKM layers such as received power, achievable rate, and effective SINR are used in this study as spatial context and as a way to validate that learned beam predictions align with site-specific propagation characteristics (Yan et al., 13 May 2026).

6. Reproducibility, extensibility, limitations, and scope of the name

SimART is organized around reproducible experiment definition. The paper states that one configuration file defines the scene, sensor sampling rates, Sionna RT settings for transmitter and receiver, Sionna SYS numerology, and the CKM grid. In the offline replay workflow, a trajectory rosbag containing poses is played, SimART nodes are launched, synchronized sensor, channel, and system outputs are published, and the entire session is recorded via rosbag. The resulting recording can be replayed in any ROS program and inspected with rviz or rqt_bag. The code is publicly available at https://github.com/guchuanv-alt/SimART, and the paper notes that the license is not specified there (Yan et al., 13 May 2026).

The platform is explicitly designed for extensibility. New environments can be introduced by swapping in any ROS-compatible simulator for aerial, ground, indoor, or maritime scenes. Sensors can be added or removed as ROS topics. Channel models can be varied through Sionna RT materials, frequencies, and arrays, while Sionna SYS exposes numerology, codebooks, and KPI computation. Algorithms can subscribe directly to synchronized multimodal topics to study beam prediction, CKM estimation, or SLAM-communication co-design. The paper’s comparative positioning emphasizes that SimART retains the strengths of robotics-only simulators while adding site-specific ray tracing, link/system evaluation, and CKMs; conversely, relative to ray-tracing-only tools it contributes platform motion, onboard sensing, and rosbag-based dataset capture (Yan et al., 13 May 2026).

The stated trade-offs are equally explicit. Mesh simplification and material annotation are required to construct propagation-ready assets. Exhaustive CKM generation is computationally heavy by design. Ray tracing can run slower than real time, although the shared simulated clock preserves alignment. Materials are assigned from a library rather than from per-site measurements. The paper demonstrates AirSim as the reference front end, but out-of-the-box adapters beyond AirSim are not detailed. Radar is not explicitly modeled in the case study, and adding new sensors requires ROS-compliant topics and, in some cases, simulator plugins (Yan et al., 13 May 2026).

The name also requires disambiguation. In the biological reaction–transport literature, the software introduced by Rangamani and collaborators is named SMART, standing for Spatial Modeling Algorithms for Reactions and Transport; the authors explicitly note that “SIMART” can appear as a misspelling in some contexts (Laughlin et al., 2023, Francis et al., 2024). Separately, an unrelated 2026 paper uses SIMART for a multimodal LLM that decomposes monolithic meshes into simulation-ready articulated assets for robotics and embodied AI (Zhang et al., 24 Mar 2026). Within the present usage, however, SimART refers specifically to the ROS-centered multimodal simulation platform for 6G ISAC described in (Yan et al., 13 May 2026).