SimART: Unified ISAC Multimodal Platform

• SimART is a unified and open multimodal simulation platform that synchronizes robotics, ray tracing, and wireless evaluation for 6G ISAC research.
• It employs modular components—Physics and Sensing, Ray Tracing, Link and System, and CKM generation—to orchestrate comprehensive ISAC experiments.
• The platform ensures precise cross-modal alignment using a shared clock, common coordinate frames, and timestamped ROS messages for reproducible datasets.

SimART is a unified and open real-world multimodal simulation platform for sixth-generation integrated sensing and communication (ISAC). It integrates mature robotics, ray tracing, and wireless evaluation engines into a single reproducible pipeline organized by a ROS backbone. Its defining mechanisms are a shared clock, a common coordinate frame, timestamped messages, and a single rosbag recording that captures the full session into one reproducible file. In this formulation, SimART is not another standalone simulator; it is an integration layer that aligns platform mobility, onboard sensing, wireless propagation, link/system-level communication metrics, and channel knowledge maps across aerial, ground, indoor, maritime, and custom ISAC scenarios (Yan et al., 13 May 2026).

1. Scope within 6G integrated sensing and communication

SimART was proposed to address a persistent tooling gap in ISAC research: robotics simulators model physics and perception but not site-specific channels, while ray tracing and link-level tools lack vehicle dynamics and onboard sensors (Yan et al., 13 May 2026). The platform is therefore motivated by the need for datasets and workflows that jointly characterize platform mobility, visual sensing and other onboard sensors, wireless propagation or channel behavior, link or system-level communication metrics, and channel knowledge maps (CKMs).

The platform’s conceptual target is cross-modal alignment under reproducible execution. Existing tools such as CARLA, AirSim, Gazebo, and Isaac Sim are strong on physics, sensing, and robotics, whereas Sionna RT/SYS, DeepMIMO, and Wireless InSite are strong on ray tracing and wireless evaluation; ns-3 is useful for network simulation but abstract for 3D geometry and multimodal sensing. SimART addresses this fragmentation by using ROS as a unifying substrate rather than replacing the underlying domain-specific engines (Yan et al., 13 May 2026).

A central architectural implication is modular reuse. The sensing front end can vary with scenario and simulator choice, while the wireless back end can remain unchanged. This suggests that SimART is best understood as a scenario-agnostic orchestration framework for ISAC experimentation rather than as a monolithic end-to-end simulator.

2. Functional modules and computational organization

SimART is organized into four functional modules: Physics and Sensing, Ray Tracing, Link and System, and CKM generator (Yan et al., 13 May 2026).

Module	Function
Physics and Sensing	Publishes platform pose and onboard sensor topics through ROS
Ray Tracing	Computes per-path propagation quantities and aggregates them into CIRs
Link and System	Converts CIRs into system-level communication outputs
CKM generator	Builds location-specific channel priors from ray tracing and system outputs

The Physics and Sensing module is defined by a ROS contract rather than a fixed simulator implementation. It must publish platform pose on the tf2 tree and onboard sensor topics including RGB images, depth images, semantic segmentation images, LiDAR point clouds, IMU, GPS, and ground-truth poses. AirSim is used as the reference implementation, but Gazebo, Isaac Sim, and CARLA can also be plugged in if they satisfy the same ROS interface (Yan et al., 13 May 2026).

The Ray Tracing module is built on Sionna RT. It loads a geometrically simplified scene matched to the visual scene’s large-scale structures, retaining buildings, roads, and major objects while removing fine details such as windows, rooftop clutter, and vegetation. It computes complex path amplitudes, delays, angles of arrival, Doppler shifts, and interaction points, then aggregates them into channel impulse responses. The module supports online mode, offline replay mode, and grid scan mode (Yan et al., 13 May 2026).

The Link and System module is built on Sionna SYS and simulates OFDM waveforms, multi-antenna chains, beamforming codebooks, and PHY/MAC behavior. Its outputs include SINR, BLER, achievable rate, and optimal beam index, where the optimal beam is defined as the beam that maximizes received SINR at each time instant. The CKM generator is native to SimART and consumes these wireless outputs to produce spatial priors for downstream ISAC algorithms (Yan et al., 13 May 2026).

3. ROS backbone, synchronization, and reproducibility

The ROS backbone is the main reason SimART functions as an integrated platform instead of a loose combination of tools (Yan et al., 13 May 2026). Three backbone mechanisms are emphasized.

First, a dedicated clock node drives all modules in ROS simulated time. This matters because ray tracing may be much slower than real time, yet sensor streams, channels, and link outputs must remain aligned. Wall-clock speed does not alter the simulated timeline, so all modules progress in lockstep under a single time reference.

Second, all important frames are placed in a shared tf2 tree rooted at a configurable world origin. The world frame, transmitter frame, receiver frame, and sensor frames therefore inhabit one coordinate system. This enables operations such as transforming LiDAR points into camera coordinates and associating a channel sample with the current transceiver pose without bespoke coordinate-handling code in each module.

Third, every ROS message carries a timestamp derived from the shared clock. This supports approximate time synchronization, multimodal alignment, and deterministic dataset generation. Because all data remain ROS-native, the entire session can be captured with a single rosbag containing sensor streams, channel impulse responses, link and system KPIs, beam indices, and coordinate frames or transformations. The resulting workflow is reproducible by construction: one configuration file specifies the scene, sensor sampling rates, Sionna RT transmitter and receiver setup, Sionna SYS numerology, and CKM grid; the trajectory rosbag player and SimART nodes are launched; ray tracing and system evaluation run in lockstep; and the full session is recorded with one rosbag record. Recorded bags can then be replayed and inspected with tools such as rqt_bag and rviz (Yan et al., 13 May 2026).

This design also formalizes the decoupling between the sensing front end and the wireless back end. Any ROS-compatible robotics simulator can supply trajectories and sensor data, while the same Sionna RT/SYS stack can be reused across disparate operational settings.

4. Spatial asset construction and electromagnetic scene abstraction

SimART provides two complementary scene-construction paths, both designed to produce a high-fidelity visual scene for the physics and sensing module and a simplified electromagnetic mesh for ray tracing, with a shared world origin embedded in both assets (Yan et al., 13 May 2026). The common coordinate frame is not a cosmetic implementation detail; it is the mechanism that ensures a position in the visual scene corresponds to the same physical location in the ray-tracing scene.

In the real-world map adaptation pipeline, SimART reconstructs a region from OpenStreetMap data. Inputs include building footprints, height attributes, road networks, and selected land-use polygons. The visual branch processes the extract with OSM2World, which extrudes buildings into 3D meshes, generates roads, and places default vegetation; the output is imported into Unreal Engine as a static level. The wireless branch loads the same OSM extract into Blender and exports it as a Mitsuba scene file for Sionna RT. Buildings, roads, and ground are assigned electromagnetic materials consistent with Sionna RT’s material model (Yan et al., 13 May 2026).

In the user-defined scene interface, controlled or non-OSM environments can be authored in tools such as RoadRunner or the Unreal Engine editor. SimART then converts the visual scene into the ray-tracing mesh with a Blender script that removes fine geometric details such as windows, railings, and decorative structures; applies quadric edge-collapse decimation; and assigns electromagnetic material types from the Sionna RT material library. The resulting mesh has 1–2 orders of magnitude fewer triangles than the original visual scene while preserving the dominant propagation geometry (Yan et al., 13 May 2026).

A plausible implication is that SimART’s scene pipeline deliberately separates geometric fidelity for visual perception from geometric sufficiency for electromagnetic propagation. The former retains appearance-critical detail, whereas the latter preserves the structures most relevant to dominant specular and diffracted paths.

5. Channel knowledge maps and the beam-prediction case study

The CKM generator creates location-specific channel priors for sensing-assisted communication and communication-assisted sensing (Yan et al., 13 May 2026). It discretizes the region of interest into a 2D ground-plane grid at a fixed receiver height. For each grid cell, a virtual receiver is placed at the cell center, ray tracing is run in grid scan mode, link and system evaluation is performed, and summary metrics are stored as CKM layers.

The output layers include total path loss, RMS delay spread, angular spread, average SINR, achievable rate, and optimal beam index. These maps function both as deterministic high-fidelity references and as ground truth for learning-based CKM estimation. The current implementation is 2D, but the paper notes that it can be extended to 3D voxel grids by sweeping multiple receiver heights (Yan et al., 13 May 2026).

The platform’s utility is demonstrated through a case study on vision- and position-aided beam prediction. A ground base station serves a UAV in an urban environment; a camera collocated with the base station observes the UAV; YOLOv8 detects the UAV in RGB images; and the detected image-plane position is combined with synchronized GPS to predict the optimal beam from a codebook. The scene is constructed through the user-defined pipeline, with paired AirSim and Sionna RT assets sharing a common coordinate frame. The hardware and communication setup uses a base station on a rooftop at 27 m, an $8 \times 8$ uniform planar array, a 3.5 GHz carrier, a 64-beam DFT codebook, and UAV altitudes from 1 m to 20 m (Yan et al., 13 May 2026).

Condition	Beam-prediction accuracy
Sunny daytime	top-1 97.86%, top-3 99.83%, top-5 100%
Rainy daytime	top-1 95.39%, top-3 99.84%, top-5 100%
Nighttime	top-1 97.92%, top-3 99.94%, top-5 100%

These results show that SimART can generate synchronized multimodal datasets with aligned visual and wireless labels and can capture realistic performance changes across environments. The paper also states that the code is publicly available at https://github.com/guchuanv-alt/SimART (Yan et al., 13 May 2026).

6. Terminology, related systems, and common ambiguity

The name “SimART” is not unique across recent arXiv literature. In the exact title-cased form, SimART denotes the ROS-based 6G ISAC platform described above (Yan et al., 13 May 2026). By contrast, SIMART in uppercase refers to a distinct 3D asset-generation framework, “Decomposing Monolithic Meshes into Sim-ready Articulated Assets via MLLM,” which converts a static 3D mesh into a simulation-ready articulated asset by jointly performing part-level decomposition and kinematic prediction with a unified MLLM and a Sparse 3D VQ-VAE (Zhang et al., 24 Mar 2026).

A further nearby source of ambiguity is ART: Articulated Reconstruction Transformer, a category-agnostic feed-forward model for reconstructing articulated 3D objects from sparse multi-state RGB images. That paper explicitly notes that it does not define a separate SimART method distinct from ART, even though “SimART” may be used informally as a shorthand or variant name in secondary discussion (Li et al., 16 Dec 2025). Other similar acronyms are also unrelated: SMART denotes Spatial Modeling Algorithms for Reactions and Transport in biological cells (Laughlin et al., 2023, Francis et al., 2024), while SimulataR denotes a desktop-based assisted-reality prototyping approach using design-blended videos (Ram et al., 27 Jan 2025).

A common misconception is therefore to treat SimART as a generic label for “simulation-ready” systems. The literature instead shows multiple unrelated expansions and research programs under closely neighboring names. In strict bibliographic usage, SimART identifies the multimodal ISAC platform centered on ROS synchronization, aligned sensing and wireless simulation, scene co-construction, and CKM generation (Yan et al., 13 May 2026).