Radio Frequency Projects Using AI and Machine Learning

MLA-formatted references with live, visible URLs • Print layout: US Letter, 1-inch margins • Generated 11 Jan. 2026

Scope note: This is a curated, representative list of publicly documented RF-focused projects and platforms where AI/ML is a core method (e.g., signal classification, spectrum sharing, neural receivers, or RAN control). It is not exhaustive, and inclusion does not imply endorsement.

  1. TorchSig. “TorchSig: A PyTorch Signal Processing Machine Learning Toolkit.” TorchSig, n.d., https://torchsig.com/. Accessed 11 Jan. 2026.

    TorchSig is an open-source RFML toolkit built around the PyTorch data pipeline, with dataset generation, augmentations, and pretrained neural models for complex I/Q signals. It is used for tasks such as signal detection, modulation recognition, and other ML workflows over digitized RF data.

  2. DeepSig. “Datasets.” DeepSig, n.d., https://www.deepsig.ai/datasets/. Accessed 11 Jan. 2026.

    DeepSig publishes the widely used RadioML modulation-recognition datasets (e.g., RML2016.10A), which underpin many ML benchmarks on I/Q samples. These datasets are commonly used to train and evaluate neural models for modulation classification and RF signal understanding.

  3. NVIDIA. “Sionna: An Open-Source Library for 6G Research.” NVIDIA Developer, 10 Dec. 2025, https://developer.nvidia.com/sionna. Accessed 11 Jan. 2026.

    Sionna is a differentiable, GPU-accelerated library for research on communication systems, designed to support gradient-based optimization and ML throughout the physical layer. It enables experimenting with neural receivers, learned components, and end-to-end training of comms pipelines.

  4. NVIDIA. “Powering AI-Native 6G Research with the NVIDIA Sionna Research Kit.” NVIDIA Developer Blog, 28 Oct. 2025, https://developer.nvidia.com/blog/powering-ai-native-6g-research-with-the-nvidia-sionna-research-kit/. Accessed 11 Jan. 2026.

    The Sionna Research Kit is positioned as an open platform for accelerating AI/ML, signal processing, and radio-propagation modeling on a unified stack. It is intended for hands-on “AI-native” wireless research workflows where training, simulation, and inference can be iterated rapidly.

  5. Defense Advanced Research Projects Agency. “Spectrum Collaboration Challenge (SC2).” DARPA, n.d., https://www.darpa.mil/research/programs/spectrum-collaboration-challenge. Accessed 11 Jan. 2026.

    DARPA’s SC2 program challenged teams to build collaborative AI/ML-driven radio systems that negotiate spectrum use in real time using software-defined radios. The program helped popularize multi-agent and reinforcement-learning approaches to dynamic spectrum sharing.

  6. Johns Hopkins Applied Physics Laboratory. “The DARPA SC2 Colosseum Test Bed.” Johns Hopkins APL Technical Digest, vol. 35, no. 1, 2019, https://secwww.jhuapl.edu/techdigest/Home/Detail?IssueID=1&Journal=J&VolumeID=35. Accessed 11 Jan. 2026.

    Colosseum is a large-scale wireless testbed built to support SC2 experiments and competition events, enabling repeatable RF experimentation at scale. Its purpose is to support AI/ML research where networks of radios adapt to interference, incumbents, and changing spectrum availability.

  7. OpenRAN Gym. “An Open Toolbox for Data Collection and Experimentation with AI in O-RAN.” OpenRAN Gym, n.d., https://openrangym.com/. Accessed 11 Jan. 2026.

    OpenRAN Gym provides an O-RAN-compliant framework for collecting RAN data and developing data-driven xApps, with an emphasis on closed-loop experimentation. It is commonly used for reinforcement learning and other ML techniques applied to scheduling, slicing, and radio resource management.

  8. O-RAN Software Community. “Welcome to O-RAN SC K Release Documentation Home.” O-RAN Software Community Documentation, n.d., https://docs.o-ran-sc.org/en/k-release/. Accessed 11 Jan. 2026.

    O-RAN SC maintains open-source implementations of RIC components and xApp infrastructure, designed to host intelligent control applications for RAN optimization. The documentation explicitly frames the non-real-time RIC as a place for analytics and model training that supplies AI/ML guidance to near-real-time control.

  9. IQT Labs. “RFClassification.” GitHub, archived 22 Feb. 2024, https://github.com/IQTLabs/RFClassification. Accessed 11 Jan. 2026.

    RFClassification is an applied ML project focused on detecting and classifying devices from RF emissions, with an example focus on drone communications. It combines feature engineering and deep learning (e.g., PyTorch) and targets eventual real-time classification pipelines.

  10. GENESYS Lab. “ORACLE RF Fingerprinting Dataset.” GENESYS, n.d., https://www.genesys-lab.org/oracle. Accessed 11 Jan. 2026.

    ORACLE is an RF fingerprinting effort that uses deep learning over I/Q samples to identify unique transmitters by learning hardware-impairment signatures. The project provides datasets intended to reproduce results and enable additional ML research on device identification and authentication.

  11. Van Hoy, Greg. “TorchDSP (gr-torchdsp).” GitHub, n.d., https://github.com/gvanhoy/gr-torchdsp. Accessed 11 Jan. 2026.

    TorchDSP is a GNU Radio out-of-tree module that integrates ML inference and DSP primitives, enabling real-time model execution inside SDR pipelines. It is designed to connect flowgraphs to inference backends for models built in frameworks such as PyTorch or TensorFlow.

  12. Suresh, Akshay, et al. “RFI-classifier.” GitHub, n.d., https://github.com/akshaysuresh1/RFI-classifier. Accessed 11 Jan. 2026.

    RFI-classifier applies supervised learning (including CNNs implemented in PyTorch) to classify radio-frequency interference morphologies in radio telescope data. It is representative of a growing set of RF/astronomy pipelines that use ML to automate interference detection and data-quality flagging at high data rates.