RadarML:
A modular ecosystem for learning on radar spectrum

What is RadarML?

Radar, in many ways, is an ideal complement to cameras for 3D perception. As a low cost¹, solid-state sensor, radar provides low range ambiguity² — as opposed to the low angular ambiguity of cameras — as well as the ability to work in the dark, through occlusions such as fog, rain, and mud³, and directly measure velocity.

	Radar	Camera	Lidar
Cost	$	$	$$$
Angular Ambiguity	High	Low	Low
Range Ambiguity	Low	High	Low

However, unlike cameras and lidar, radar data are difficult to collect and unintuitive to interpret. Especially in the case of raw I/Q data, few tools and datasets exist, creating a high barrier to entry for research. Our goal is to fill this gap by providing high-quality, modular, and fully open-source data collection tools, processing pipelines, datasets, and research frameworks to enable learning on radar spectrum.

Active Projects

These projects are in active development, and will continue to receive new (potentially substantial) features and updates. All of these projects are actively supported, and we welcome contributions from the community.

• nrdk

  

  neural radar development kit for deep learning on multimodal radar data

• xwr

  

  python interface for collecting raw time signal data from TI mmWave radars

• xwr_ros

  

  ros2 node and data types for xwr

• firmware

  

  modified firmware for TI mmWave radar development boards removing key limitations with the default firmware

• abstract_dataloader

  

  abstract interface for composable dataloaders and preprocessing pipelines

Completed Projects

These projects are considered to be feature-complete, and will only receive bug fixes or other minor updates. Community contributions are still welcome, but we generally do not expect to make substantial additions or changes.

• red-rover

  

  a system for collecting and processing mmWave Radar, camera, and Lidar data

• i/q-1m

  

  29 hours of radar (time signal), lidar, and camera across indoor, outdoor, and bike-mounted settings

Upcoming Projects

• mmwcas

  

  data collection and processing for the TI MMWCAS cascaded imaging radar

• tartan*

  

  an upcoming dataset with radar, lidar, stereo RGB, thermal, and event cameras

Other Resources

• 18-848 RadarML @ CMU

  

  Special Topics in Embedded Systems: Machine Learning for Radar, Fall 2026

• Radar Cheatsheet

  

  FMCW radar cheatsheet and quick reference manual

• open-svo2

  

  Reverse engineering of the StereoLabs SVO2 File Format

---

1. Single-chip mmWave radars typically cost $10-50, compared to $1000+ for lidar. ↩

2. Radars provide constant range resolution regardless of distance, while cameras can only measure depth via triangulation (either as a stereo pair or via structure-from-motion), which degrades with distance (relative to baseline). ↩

3. Unlike cameras and Lidars, mmWave radars can see through sensor occlusions such as mud or paint, and are robust to atmospheric occlusion such as fog and rain. By exploiting multipath, radars can also see around objects to some extent. ↩