SEAM Studio GitHub ↗
Hanyang University campus digital twin rendered in SEAM Studio with real aerial textures
NEW v0.1.6 channel-dataset npz export from the GUI · reproducible scene hashes · path-buffer control — pip install -U seam-studio
Scene-to-Electromagnetic Authoring and Mapping

SEAM Studio

A local-first, AI-assisted RF digital twin workbench. In one textured 3D scene, every mesh carries two material bindings — visual (PBR) and electromagnetic (RF). The canonical scene compiles into a Sionna RT projection, and ray paths and radio maps come back as overlays in the same viewport. No GPU, no Sionna, no LLM required.

View on GitHub User guides Quickstart
HYU campus · drone/aerial-textured twin, imported & rendered in SEAM Studio
SEAM Studio full app — 3D viewport with the Visual/RF Materials/Validation/AI Assist/Results modes and the communication-metrics dashboard
SEAM Studio, in one view — the 3D scene viewport, the Visual / RF Materials / Validation / AI Assist / Results modes, and a live communication-metrics dashboard in a single workbench.

What it does

From scene import to radio propagation, on one screen

Import Mitsuba XML, zip bundles or an OpenStreetMap rectangle; assign and validate RF materials; run 28 GHz ray tracing; and paint the results back onto the same scene.

Author

Scene & material authoring

Texture-embedded GLB viewer (Blender-style navigation), per-prim RF materials, physical splitting.

  • zip bundle / OSM map-drag import
  • material & connectivity mesh splits + undo
  • ITU band guardrails & validation lints
Simulate

Sionna RT projection

Canonical scene → Mitsuba XML auto-compile, with automatic recompile on staleness.

  • paths · radio maps · SINR · beamforming · channel (CFR)
  • multi-UE trajectories + terrain-following drape
  • ML ground-truth dataset export
SEAM-Agent

Retrieval-augmented local AI material assignment

A local LLM/VLM harness observes buildings in multi-view, augments with web evidence, and proposes materials.

  • triangle-id back-projection → per-face segments
  • observable activity trace + evidence cards
  • confidence · provenance · user approval

Pipeline

Unsegmented mesh → Sionna-ready RF scene

1 · Import

Mitsuba XML / zip / OSM. Textures are stored twice: viewer GLB + full-resolution AI evidence.

2 · Segment

Per-face splits from heuristics, a local VLM, or an uploaded SAM2 mask.

glassconcretemetalground
3 · Assign

SEAM-Agent proposals + user review. Every decision recorded with provenance.

4 · Compile

Per-material PLY groups + clean Mitsuba XML with ITU bsdfs, generated automatically.

5 · Solve

Sionna RT (or Mock) paths & radio maps → overlaid on the same viewport.

In action

The tool in operation

Sionna RT ray tracing on the campus twin — 39 paths between one TX and two RX devices
Sionna RT ray tracing — 39 paths (LOS cyan · reflections magenta) between one TX and two RXs on the campus twin, solved by the real Sionna backend.
Top-down view of the textured campus twin with continuous terrain
Campus twin, top-down — aerial textures on continuous terrain, with view snaps (1/3/7) and an infinite grid.
Sample demo scene with TX and RX devices
Sample Demo — TX/RX placement and height-above-surface (AGL) editing; the tutorial scene runs on the Mock backend alone.
Drone-mapped FTC building multi-view capture used by SEAM-Agent
Drone-mapped FTC building — the orthographic multi-view capture that SEAM-Agent sees.

Capabilities

A closer look at the features

Under the three pillars sit the tools you run experiments and write papers with — from link analysis to dataset generation.

Channel analysis & live tuning

Link budget, CIR/CFR, K-factor, RMS delay spread, and empirical path-loss models (incl. 3GPP 38.901) compared against RT. Change frequency, bandwidth, TX power, noise figure or SCS and the analysis re-runs instantly, including TS 38.215-style RSRP/RSSI/RSRQ.

Multi-TX interference & SINR

Ray-traced received powers from all non-serving TXs sum into co-channel interference: SINR = S/(I+N). Selectable serving cell, SINR radio maps, and a per-cell serving-TX map.

MIMO beamforming

Codebook sweep / TX-MRT / SVD beamforming gains computed on the actual ray channel (~12/24 dB validated at 4×4).

Mesh radio maps + region refinement

Coverage painted per triangle on real surfaces — walls, floors, roads — instead of a horizontal plane, with region-of-interest re-solves at finer cell sizes.

Trajectory RF metrics & terrain following

Sample RSS, path gain, RMS delay and SINR along UE trajectories. Trajectories drape onto terrain and rooftops instead of tunneling through hills, and devices are placed by height-above-surface (AGL).

Material segmentation & splitting

Split a monolithic building mesh into per-material faces from a texture mask (color heuristic / local VLM / uploaded SAM2 mask), or split merged meshes into connected components. Every split keeps a GLB backup and is undoable.

Measurement-driven calibration

Import measured link CSVs to grid-fit material parameters, tell visually identical materials apart from measured path gains, and quantify how much an assignment matters via NMSE and capacity deltas.

Metrics dashboard & paper-ready export

Link KPIs and CIR/CFR/Doppler/path-loss charts in a white-background serif paper style — every chart has built-in PNG/SVG/CSV export. Viewport capture and offline Mitsuba renders included.

Datasets & AODT interop

Generate ML ground-truth datasets (npz + metadata), export the AODT RFData contract, and import NVIDIA AODT parquet results into the same schema.

Natural-language rules & validation explanations

Turn “windows are glass, concrete walls are itu_concrete” into reviewable assignment rules applied in bulk, and get validation warnings explained in plain language with one-click actions.

Result reproducibility & live events

Every result is stamped with scene/assignment/config hashes plus a config snapshot, so stale results get badges; compile and simulation progress streams over WebSocket.

Live sync POST /live/state

Inject real-world positions from GPS, mocap or logs into the loaded scene — the viewer follows in real time and can re-simulate immediately for a measure → sync → predict loop.

SEAM-Agent

“This building is the Hanyang FTC” — one line is enough

Give the agent a hint and it retrieves real exterior photos from the web, while a local VLM extracts material cues and fuses them with multi-view observations. The whole run is an observable activity trace, and nothing is applied without your approval.

Real exterior photo of the Hanyang FTC building retrieved from the web by the agent The drone-mapped FTC mesh after per-segment RF material assignment — glass curtain wall shown in blue
Left: the real FTC exterior photo the agent retrieved (Modamoda, CC BY 3.0, Wikimedia Commons) · Right: the same building's drone-mapped mesh after per-segment assignment, in RF material color mode (glass curtain wall = blue, concrete = grey)
[done] web_search "Hanyang University Fusion Technology Center FTC exterior"
[done] 6 photos retrieved · Wikimedia + hyu.wiki
[done] VLM claim: "glass curtain wall, white concrete structure"
[done] 4 views analyzed → face votes → segments
[review] curtain_wall_glass → itu_glass (0.83) · roof → itu_concrete (0.98)

Quickstart

Just want to run it? — pip

Any OS · no Node.js, no checkout
pip install seam-studio
seam-studio   # serves + opens http://127.0.0.1:8000
What you get
# pre-built UI bundled in the wheel — Node.js not needed
# sionna-rt ray tracing installs as a base dependency
# first run seeds a Sample Demo project in ~/.seam/projects

Developing or using the repo examples? — from source

Windows · PowerShell
git clone https://github.com/jaewoo4200/SEAM.git; cd SEAM
powershell -ExecutionPolicy Bypass -File scripts\install.ps1
powershell -ExecutionPolicy Bypass -File scripts\start.ps1
# → http://localhost:5173
Linux · macOS
git clone https://github.com/jaewoo4200/SEAM.git && cd SEAM
bash scripts/install.sh
bash scripts/start.sh   # → http://localhost:5173

Prerequisites: Python 3.11–3.14; the source route also needs Node.js 20+ on your PATH (Windows: winget install OpenJS.NodeJS.LTS, then open a new terminal). Everything else is handled for you, including the real sionna-rt ray-tracing engine — it is a base dependency, and the app falls back to the bundled Mock backend whenever Sionna cannot load. An NVIDIA GPU and a local LLM (LM Studio/Ollama) are optional upgrades. See INSTALL.md · TUTORIAL.md · or start with the illustrated user guides.