Roadmap

Phased delivery plan. Each milestone is a vertical slice: a working dataset that exercises every layer of the framework at the appropriate fidelity for that phase.

Note

Effort estimates assume two engineers full-time. The relative ordering and dependency structure matter more than the absolute dates.

v0: Single-signal narrowband + caption text

Goal. Validate the two-phase pipeline end-to-end on the simplest variant: one emitter, narrowband, AWGN-only, single-RX, one annotation type.

Datasets. narrowband_classifier_baseline_md (100K samples, RadioML 2018-aligned).

Features. TorchSig adapter (24 RadioML modulations) · TorchSigImpairments level 1 · single-emitter “trivial scene” · class label only (no bbox/seg) · caption annotator (Gemini Flash) · Zarr + WebDataset · LocalRunner only · statistics + PAES audits.

Acceptance. Caption PAES ≥ 0.85; schema compliance ≥ 99.5 %; AMC head trained on it matches RadioML 2018 difficulty curve within 5 %.

Out of scope. Multi-emitter; bbox/seg; Sionna; fingerprint; multi-RX; verifier; HIL; HDF5; Spark; >1 M samples.

Effort. ~6 engineering weeks.


v0.5: Multi-emitter with TorchSig + bbox labels

Goal. Multi-emitter wideband generation with bbox + segmentation labels at scale on cloud orchestration. Functional parity with TorchSig WidebandSig53 plus our additions.

Datasets. wideband_detection_baseline_md/lg (100K / 1M); narrowband_classifier_baseline_xl (10M, scaling-law proof point).

Features. All 60+ comms emitters · TorchSigImpairments 0/1/2 + AWGN + SionnaTDL · default scene composer · full label schema v1.0 · Zarr + WebDataset + HDF5 TorchSig interop · Managed Spark serverless with idempotent shard naming · cross-modality consistency audit.

Acceptance. Cross-modality consistency 100 %; HDF5 round-trip byte-exact for TorchSig fields; statistics audit within 1σ of WidebandSig53; idempotent re-run leaves no duplicates.

Out of scope. Sionna channels; fingerprint; radar/drone/IoT/ADS-B/cellular; LLM annotations; HIL; multi-RX.

Effort. ~10 engineering weeks.


v1: Heterogeneous zoo + Sionna channels + LLM Q&A

Goal. Full heterogeneous emitter zoo, ray-traced channels, and the complete annotation suite. The first release that delivers on the framework’s headline promise.

Datasets. heterogeneous_dense_2_4ghz, contested_ew_lband, device_fingerprint_lora_25, multimodal_drone_sim, each at _md and _lg.

Features. Radar (pulse, FMCW, LFM) · drone (OcuSync, Lightbridge) · IoT (LoRa, BLE, Zigbee, Wi-Fi) · ADS-B · LTE + 5G NR · per-device fingerprint module · Sionna RT (ray-traced multipath) + Sionna PHY (TDL/CDL) · all 5 annotation types (caption, QA, reasoning, scene_report, contrastive) · verifier subset (5 % via Sonnet) · Vertex AI Batch Prediction · taxonomy unifier across 9 source vocabularies.

Acceptance. PAES ≥ 0.80 on all annotation types · Hallucination Count ≤ 0.5 / sample on verifier subset · Sionna RT scenes reproducible by seed · cross-dataset taxonomy lookup 100 %.

Out of scope. HIL validation; multi-RX in production presets; sim2real benchmark.

Effort. ~16 engineering weeks.


v1.5: HIL validation + sim2real benchmark

Goal. Ground the synthetic dataset against real-radio captures. Without this, no claim about realism survives review.

Deliverables. USRP/HackRF cabled-loopback + OTA capture harness · 1K-sample HIL subset across 10 emitter classes · automated re-labeler · sim2real benchmark report (per-class accuracy delta, channel statistics delta) · multi-RX scene presets (4-element ULA).

Acceptance. Sim2real accuracy delta ≤ 5 pp on classification heads · channel-statistics KL-divergence ≤ 0.1 against captured baseline · multi-RX presets reproducible.

Effort. ~12 engineering weeks (much of it lab time, not coding).


v2: CLIP-for-RF released dataset

Goal. Public release. A permissively-licensed RF + text dataset and a CLIP-for-RF baseline trained on it.

Deliverables. Filtered subset of v1 datasets restricted to permissively-licensed source classes · CLIP-for-RF training recipe · pretrained encoder weights · benchmark suite + leaderboard.

Acceptance. External users can reproduce baseline numbers from the published recipe within 2 % · dataset license review cleared by legal · benchmark accepted at one venue.

Effort. ~10 engineering weeks of work after v1.5.


Cross-project dependencies

This framework feeds the rf-foundation-models project. Each milestone unblocks one of theirs:

Our milestone

Unblocks

v0

Their AMC head pretraining

v0.5

Their wideband detection / multi-task heads

v1

Their RF-LLM Q&A and reasoning evaluation

v1.5

Their sim-to-real benchmark report

v2

Joint public release