Our vision
Multimodal design
One ViT backbone learns a shared representation for IQ streams, spectrograms, CSI, and CIR.
Multitask support
Jointly supports sensing, communication, and localization tasks.
Benchmarks + artifacts
Linked datasets, protocols, reproducible benchmarks, weights, and code.
Benchmarks at a glance
Compare performance across datasets and fine-tuning regimes.
How to read
Regimes: LP = linear probe, FT2 = partial fine-tuning, LoRA = low-rank adaptation. Positioning tasks are displayed as 1/(1 + error) so higher is better.
Why WavesFM
Masked wireless modeling
Self-supervised pretraining to learn a shared representation across modalities.
Downstream evaluations
Evaluate on a range of tasks: human activity sensing, signal and modulation classification, NR and UWB positioning, beam prediction, and more.
Finetuning regimes
LP, FT2, and LoRA finetuning.
Reproducibility
Docs, recipes, and versioned releases.
Paper
Main reference for the current model and benchmark suite.