Literature Review

The framework is built on top of and informed by a substantial body of prior art. This page covers the headline references; gap analysis and how each citation flows into a design decision are in Design Decisions.

Headline references

Synthetic RF data generation

  • TorchSig (Boegner et al., 2022; v2.1.x at writing): BaseSignalGenerator, Transform, HDF5 I/O, 60+ comms generators. The substrate we wrap.

  • Sionna (Hoydis et al., 2022): link-level simulation, including sionna.rt (ray tracing) and sionna.phy.channel (TDL/CDL). Channel backend.

  • WidebandSig53: TorchSig’s wideband multi-emitter dataset; the closest existing dataset to what we produce.

RF foundation models and multimodal

  • RF-GPT: RF + text pairing via deterministic templates plus LLM rewrite. Annotation pattern.

  • WavCaps (Mei et al., 2024): three-stage audio captioning pipeline (template, bulk LLM, verifier). Pipeline pattern.

  • Qwen-Audio (Chu et al., 2023): hierarchical-tag conditioning to unify heterogeneous audio label spaces. Taxonomy pattern.

  • CLIP (Radford et al., 2021): contrastive image-text pretraining. The downstream objective our v2 targets.

Wideband detection and segmentation

  • WRIST (2021): spectrogram-as-image YOLO/DETR detection. Bbox conventions.

  • ZoomSpec (2026): physics-guided STFT preprocessing for RF detection. STFT default params.

  • Aboelazm 2025: multi-emitter detection benchmark; scene density realism.

Source datasets we draw from

RadioML 2016.10a / 2018.01a, Sig53, HisarMod 2019.1, DroneRF, RFUAV, CardRF, WiSig, SMoRFFI, Daytona ADS-B, NIST CBRS Radar.

Sim-to-real and validation

  • PAES (Physical Attribute Extraction Score): caption quality metric grounded in verifiable physical attributes.

  • IQFM / LSMs: large signal models; the broader research direction this framework feeds.

Gaps this framework fills

  1. Heterogeneous emitter zoo under one interface. TorchSig is comms-only. RadarSimPy, gr-lora_sdr, pyModeS, etc. each ship their own SDK and label conventions. We unify them under BaseEmitter.

  2. Pluggable channel realism. No existing pipeline lets you swap stochastic TDL ↔ ray-traced multipath ↔ hardware impairments by changing one config flag.

  3. Joint labels for the same record. Existing datasets ship bbox or segmentation or per-emitter metadata; no canonical store has all three plus text annotations on the same sample.

  4. Inference-grounded annotations with hallucination control. RF-GPT and WavCaps don’t address hallucination explicitly; we add a verifier subset and JSON-constrained decoding.

  5. Sim-to-real validation harness. Most synthetic RF datasets are pure simulation; ours ships an HIL playback harness on day one.