Why the radio is a good place for ML
Classical radio algorithms are hand-designed from models — a channel estimator assumes a certain channel, a codebook assumes certain antenna geometry. Reality is messier than the models. Machine learning offers algorithms that adapt to the actual deployment from data. Three jobs are especially ripe, because they are high-volume, pattern-rich, and expensive to do optimally by hand: channel state feedback, beam management, and positioning.
ITU-R M.2160 names Artificial Intelligence and Communication as a headline 6G usage scenario — the network as a distributed compute/AI fabric. That makes AI a requirement-level driver for 6G, not an optional extra.
itu.int — IMT-2030 2023-11The 6G AI/ML-for-air-interface direction builds directly on completed and ongoing 3GPP work (the Rel-18 study captured in TR 38.843, continued in Rel-19) covering CSI feedback/compression, beam management and positioning. These are the proving grounds; how much carries into normative 6G is a Release 21 question.
wirelessbrew.com 2026-03 secondaryThe hard part: two-sided models
Channel-state-information (CSI) feedback is the flagship example. Today a phone measures the channel and reports a compressed summary so the tower can beamform. Replace the compressor with a neural encoder on the phone and a neural decoder at the tower and you can shrink the report dramatically. But now two different vendors' models — one in the phone, one in the base station — must agree, like two halves of a codec built by strangers. That "two-sided model" coordination (training, versioning, interoperability) is the genuinely novel standardisation problem 6G must solve.
The functions ML aims to improve are not new — 5G NR already has CSI reporting and a beam-management framework (SSB/CSI-RS beam sweeping, reporting, indication). Understanding the 5G versions is the fastest way to see precisely what ML changes and what stays the same.
ranbits — Foundations (5G beam sweep) 2025| Use case | 5G NR approach (foundation) | AI/ML candidate approach | Key challenge | 3GPP study | Status |
|---|---|---|---|---|---|
| CSI feedback compression | Codebook-based Type I / Type II reporting; fixed compression | Neural encoder (UE) + decoder (gNB) — data-driven; latent as small as 16 values | Two-sided model: vendor interoperability, versioning, retraining | TR 38.843 (Rel-18) — feasibility confirmed | candidate |
| Beam management | SSB / CSI-RS sweep → UE reports RSRP → gNB selects TCI state | ML predicts best beam from history / position, reducing sweep overhead | Generalisation across environments; beam-failure latency | TR 38.843 (Rel-18) — studied | candidate |
| Positioning | OTDOA / UL-TDOA (Rel-16); DL-AoD / UL-AoA (Rel-17) | ML exploits multipath fingerprinting for sub-meter NLOS accuracy | Training data per deployment; NLOS vs LOS classification | Rel-18/19 study | candidate |
| Channel estimation | Pilot-based MMSE / LS; interpolation across REs | Model-aided / deep learning CE — better in sparse-pilot scenarios | Complexity vs gain trade-off; low-latency inference on UE | 6G study item | candidate |
Where this connects
AI rides on the physical layer (it learns the waveform's channel) and is pulled by the requirements (AI & Communication is a named scenario). It also overlaps sensing, where ML interprets the echoes the radio collects.