SEAM Studio

Unified RF–visual scene authoring, RF material assignment, and Sionna RT digital-twin simulation — in one local workbench.

SEAM Studio pairs a FastAPI backend with a bundled React/three.js frontend: author a 3D scene, bind ITU/custom RF materials to its surfaces (by hand, by rules, or with a local-LLM agent), then ray-trace paths, radio maps, UE/UAV trajectories, Doppler, beamforming, handover and radar sensing / ISAC — all persisted as reproducible result sets.

Install

pip install seam-studio

This installs the real ray-tracing engine (sionna-rt, Dr.Jit/Mitsuba) by default. It runs on an NVIDIA GPU (CUDA). Without one, Dr.Jit's CPU backend needs LLVM installed separately (Windows: the official LLVM installer, then set DRJIT_LIBLLVM_PATH to its LLVM-C.dll; Linux: LLVM 18+, e.g. apt install libllvm18 on Ubuntu 24.04, from apt.llvm.org on Ubuntu 22.04 / Debian 12 whose default LLVM 14 is too old; macOS: Xcode CLT or brew install llvm). When neither backend works, the app reports Sionna as unavailable with the reason (/api/health) and auto runs its deterministic Mock engine instead. Details: INSTALL.md.

Optional extras: pip install "seam-studio[parquet]" adds pyarrow for AODT parquet import/export and parquet sensing datasets.

Quickstart

seam-studio            # starts on http://127.0.0.1:8000 and opens the browser

The first run creates a Sample Demo project (toy urban scene with a rooftop TX and its co-located sensing RX, a street RX, car and pedestrian actors, and a drone sensing target) under ~/.seam/projects, so you can press Simulate paths immediately.

seam-studio --port 9000            # different port
seam-studio --project-root D:\twins  # keep projects elsewhere
seam-studio --no-browser

Highlights

• Scene → RF binding: import Mitsuba XML / scene bundles / OpenStreetMap extrusions; assign RF materials per surface with validation and provenance.
• Simulation: paths, planar & mesh radio maps, multi-TX SINR/RSRP/RSRQ, MIMO beamforming, UE/UAV trajectories with per-step handover (3GPP A3), Doppler spectrograms, ML ground-truth dataset export (NPZ).
• Radar sensing & ISAC: RCS targets (TR 38.901 / constant), echo solves with Doppler, per-frame multistatic fusion and EKF tracking, Swerling Pd/Pfa detectors with Monte Carlo, ISAC beam trade-off, sensing coverage maps, labeled sensing dataset export (npz/csv/parquet).
• AI assist: local LLM/VLM material suggestion agent (Ollama / LM Studio), natural-language assignment rules, validation explains.
• Reproducibility: every run persisted with config snapshots + content hashes; measurement import and calibration against real logs; NVIDIA AODT parquet import/export.

Links

• Repository & docs: https://github.com/jaewoo4200/SEAM
• Website: https://jaewoo4200.github.io/SEAM/

Apache-2.0 © Jaewoo Lee