From raw returns to actionable products
6G sensing produces high-dimensional data: range-Doppler maps, angle spectra, multi-static delay profiles and time series of complex baseband samples. Classical signal processing extracts detections and tracks through matched filtering, constant-false-alarm-rate detectors and multi-target trackers. Machine learning augments and, in some cases, replaces stages of this pipeline when the environment is complex, the clutter is non-stationary or the desired output is semantic rather than purely geometric.
Typical AI targets include detection under heavy clutter, classification of target type or activity, semantic segmentation of occupancy maps, and prediction of future trajectories. Models range from convolutional networks on range-Doppler images to graph networks on multi-static measurements and recurrent or transformer architectures on temporal sequences.
Edge and on-device constraints
Many sensing applications—automotive safety, industrial safety zones, fall detection—require low latency and cannot tolerate round-trips to a distant cloud. Models must therefore run on radio units, edge servers or even user equipment under strict power, thermal and memory limits. Model compression, quantization, pruning and architecture search are essential. Hardware accelerators integrated into baseband system-on-chips or attached neural processing units provide the necessary throughput per watt.
Online adaptation is another practical requirement. A model trained in one factory or city block may degrade when the clutter distribution changes. Lightweight fine-tuning, domain adaptation or continual learning techniques that respect privacy and computational budgets are active research topics.
Data, labels and simulation
Supervised learning needs labels. Collecting labeled 6G sensing data at scale is expensive; privacy constraints further limit sharing of real measurements that contain human activity. High-fidelity electromagnetic simulation, ray-tracing with material libraries and generative models help create synthetic training sets. Hybrid training that mixes synthetic and real data, together with self-supervised pre-training on unlabeled real returns, reduces reliance on exhaustive manual labeling.
Evaluation must go beyond accuracy. Latency, energy per inference, robustness to distribution shift and calibration of confidence scores all matter when models feed safety or operational decisions.
Integration with classical processing
Pure end-to-end neural pipelines are attractive in research but rare in production systems that must provide certifiable false-alarm rates or explainable outputs. Hybrid designs keep classical front-ends for pulse compression and basic detection, then apply learned models for classification, association or semantic mapping. Learned components can also improve classical estimators—for example, by predicting residual clutter or by refining direction-of-arrival estimates under array imperfections.
Governance and lifecycle
Models that influence safety or privacy decisions require version control, testing under representative conditions, monitoring for drift and clear rollback procedures. When sensing data leave the edge, access control and purpose limitation still apply. Operators and vendors need processes that treat AI models as part of the network software lifecycle, not as disposable experiments.
AI does not remove the need for good waveform design, calibration and resource management. It amplifies the value of the measurements those systems produce by turning them into reliable, low-latency products that applications can consume.
