Concepts

Audience. Engineers implementing or extending the framework, plugin authors adding new emitters or channels, and downstream data consumers who want to understand how records are structured.

Before you start. Complete Getting Started to have the install and quickstart behind you. Familiarity with Python and basic DSP (complex baseband, sample rate, SNR) is assumed.

What this section is not. This section is not a step-by-step how-to guide. For task recipes (“how do I generate a dense urban dataset?”), see How-to guides.

---

Concept pages explain the mental model for the RF data-generation framework. Dense schemas, class signatures, storage layouts, and exact option tables live in Reference.

Core Ideas

• Compose mature tools before writing custom framework code. Sionna, TorchSig, NumPy, SciPy, PyTorch, Zarr, WebDataset, HDF5, SigMF, Hydra-style config, and Pydantic-style validation are default building blocks where they fit.

• Keep stable contracts at framework boundaries. Emitters, channels, scenes, labels, annotations, storage, and executors are swappable because they exchange framework records with documented metadata and lifecycle rules.

• Keep adjacent concepts separate. An emitter is not a scene composer; a channel is not a receiver frontend; a labeler is not an annotator; a storage writer is not a training adapter.

• Make claims testable. Determinism, label correctness, schema compatibility, and annotation grounding should point to validation checks in Reference / Test Contracts.

Pages

The framework is six layers, plus shared concepts that describe the records and coordinate systems moving through those layers.

Layer	What it does	See
Architecture	The pipeline as a whole: data flow, sample lifecycle, extension model	Architecture
Core Types	Stable data contracts: Signal, Record, metadata, bboxes	Core Types
Coordinate Systems	Scene center frequency, emitter carrier offsets, scene-relative time	Coordinate Systems
Records, Receivers, and Assets	Multi-TX baseline, multi-RX record layout, and content-addressed Sionna RT assets	Records, Receivers, and Assets
Emitters	Per-family signal generators (comms, radar, drone, IoT, ADS-B, cellular)	Emitters
Channels	Four-group transformation pipeline: TX impairments, channel, RX capture, RX hardware	Channels
Scenes	Multi-emitter composition: density, SNR mix, event timing, multi-RX	Scenes
Labels	Joint format: bbox + segmentation + per-emitter metadata	Labels
Annotations	Inference-grounded captions, Q&A, reasoning chains, scene reports	Annotations
Storage	Backends for persisting records: Zarr canonical, WebDataset training shards, HDF5 export, SigMF input	Storage
Evidence and Release Qualification	Hash-linked, signed records that make a dataset’s claims independently checkable	Evidence and Release Qualification
Consumer Access	Sampler, collator, and PyTorch adapter that wrap the storage read API	Consumer Access
Signal Catalog	Supported emitter families and coverage status	Signal Catalog

For the why behind these choices, see Background / Design Decisions.