Source-linked AI summary
ColoRadar: The Direct 3D Millimeter Wave Radar Dataset
Andrew Kramer, Kyle Harlow, Christopher Williams, Christoffer Heckman
TL;DR
Public datasets with dense, high-resolution radar data for 3D odometry and mapping are scarce. ColoRadar provides a multisensor dataset with several radar data forms and supporting sensor measurements for 6DoF state estimation and mapping. Its diverse 3D environments and radar formats offer data for developing robotic perception methods.
Problem
Public robotics datasets with radar data are scarce, and existing radar datasets lack raw ADC samples needed to develop alternative signal-processing and target-detection methods.
Source: Motivation and Contribution · p. 1Method
ColoRadar provides raw ADC samples, dense radar heatmaps, and onboard-derived point targets alongside lidar, IMU, and groundtruth pose information.
Source: Motivation and Contribution · pp. 1–2, Data Description · p. 4Results
The dataset combines radar measurements from multiple sensors and formats with slow and rapid 6DoF motion across indoor, outdoor, and subterranean environments.
Source: Related Work · p. 2, Summary and Future Work · p. 7Takeaways & Limitations
The dataset supports developing signal-processing, target-detection, state-estimation, and mapping methods for small agile robots in diverse 3D environments.
Source: Related Work · p. 2, Data Description · p. 4
Abstract
from arXiv · showhide
Millimeter wave radar is becoming increasingly popular as a sensing modality for robotic mapping and state estimation. However, there are very few publicly available datasets that include dense, high-resolution millimeter wave radar scans and there are none focused on 3D odometry and mapping. In this paper we present a solution to that problem. The ColoRadar dataset includes 3 different forms of dense, high-resolution radar data from 2 FMCW radar sensors as well as 3D lidar, IMU, and highly accurate groundtruth for the sensor rig's pose over approximately 2 hours of data collection in highly diverse 3D environments.
Introduction
Robotic localization and mapping can rely on radar in visually degraded environments, but sensor and data-access constraints—and the lack of suitable public datasets—hinder method development. ColoRadar addresses this gap with a multisensor dataset for 6DoF radar-based state estimation and mapping, offering radar data at multiple processing levels.
Motivation and Contribution
- Visual and lidar estimates are severely handicapped in visually degraded environments such as darkness, fog, and smoke.The paper notes these conditions in settings of interest including disaster response and subterranean exploration.Source: Motivation and Contribution · p. 1
- Millimeter wave radar operates without ambient light and is unaffected by airborne smoke and fog, while automotive SoC sensors are lightweight and low-power.The paper presents these properties as attractive for perception in degraded environments and for platforms with limited payload and power, such as MAVs.Source: Motivation and Contribution · p. 1
- Public robotics radar datasets with the data needed to develop radar processing methods are unavailable, while streaming full data from most sensors is difficult and time-consuming.The cited barriers include sensors that restrict access to full data streams and those that are not designed for real-time streaming.Source: Motivation and Contribution · p. 1
- The authors present a multisensor dataset designed for 6DoF radar-based state estimation and mapping, including dense radar data, lidar, IMU, and groundtruth poses.Radar is supplied as raw ADC samples, dense heatmaps, and onboard-processed point targets.Source: Motivation and Contribution · pp. 1–2
Related Work
Existing radar datasets have limitations in motion, environment, sensor type, or access to raw radar data. ColoRadar combines 6DoF motion, multiple 4D radar data forms, and diverse environments to support methods for agile robots in 3D settings.
Related Work
- Oxford Radar RobotCar and nuScenes focus on road vehicles with motion largely limited to a 2D horizontal plane, while the cited Euroc MAV and UZH FPV datasets are visual-inertial rather than radar datasets.The Chilean underground mine dataset also uses 2D scanning radar and is mostly confined to 2D movement.Source: Related Work · p. 2
- Existing radar datasets lack raw ADC samples, leaving users dependent on onboard signal-processing and target-detection algorithms that the authors describe as often ill-suited to robotics tasks.The cited contrast includes dense 2D horizontal scans from the NavTech sensor and road-vehicle-focused datasets.Source: Related Work · p. 2
- ColoRadar includes slow and rapid 6DoF motion, raw ADC and dense data from multiple 4D radar sensors, and indoor, outdoor, and subterranean environments.The authors connect this breadth to developing state-estimation and mapping methods for small agile robots in diverse 3D environments.Source: Related Work · p. 2
Sensor Platform
The sensor rig combines two FMCW radar sensors with lidar and an IMU. Its radar designs trade angular resolution against framerate, data rate, and power draw.
Sensor Platform
- Figure 1 labels the sensors in the rig and marks each sensor's local coordinate frame.The platform is shown in Figure 1.Source: Figure 1 · p. 2, Sensor Platform · p. 2
- The platform uses a cascaded TI MMWCAS-RF-EVM radar and a TI AWR1843BOOST-EVM single-chip radar with a DCA1000-EVM for raw-data capture.The rig also includes an Ouster OS1 lidar and a Lord Microstrain 3DM-GX5-25 IMU.Source: Sensor Platform · p. 2
- The cascaded radar provides far better angular resolution than the single-chip sensor, at the cost of lower framerate, higher data rate, and higher power draw.Both sensors use an FMCW waveform, so the same signal-processing and target-detection algorithms may be used for either.Source: Sensor Platform · p. 2
Radar Signal Processing
The paper describes how FMCW MIMO radar measurements are transformed into range, velocity, and angular heatmaps. Processing spans chirp-based range and velocity estimation, virtual antenna arrays, and angle-of-arrival transforms; onboard target detection can reduce measurement size but discards information and introduces false alarms.
Range-Doppler Velocity Processing
- Fast-time and slow-time FFTs convert chirp samples into received power over range and range rate for each virtual antenna.Samples within a chirp form fast time; repeated chirps are stacked as a frame, whose height is slow time.Source: Range-Doppler Velocity Processing · p. 3
- Each MIMO transmit-receive antenna pair acts as a virtual antenna at the pair's centroid, with spatially varying phase supporting angle-of-arrival processing.The transmitted signals are multiplexed in time, yielding NT NR unique received signals per frame.Source: Range-Doppler Velocity Processing · p. 3
Angle Of Arrival Processing
- Range-Doppler outputs are stacked into a complex data cube, and an FFT across virtual antennas yields a heatmap over range, Doppler velocity, and azimuth.The cube contains samples for multiple virtual antennas.Source: Angle Of Arrival Processing · p. 3, Figure 4 · p. 4
- Using virtual antenna elements distributed in two dimensions extends the heatmap with elevation, producing range, Doppler velocity, azimuth, and elevation dimensions.The two-dimensional antenna-array processing applies FFTs across both array dimensions.Source: Angle Of Arrival Processing · p. 3
- CFAR detection returns prominent heatmap peaks as point targets, reducing measurement size while discarding potentially valuable information and introducing false alarms.The paper describes false alarms as a portion of detected targets.Source: Angle Of Arrival Processing · pp. 3–4
Data Description
The dataset spans diverse indoor, outdoor, mine, and motion-capture environments and provides multiple radar processing levels alongside lidar, IMU, and pose data. Its passages describe binary storage and per-sensor formats, including groundtruth and Vicon poses.
Data Description
- Sequences cover indoor and outdoor built environments, a dynamic creek path, open and constrained mine areas, and an obstacle-filled motion-capture space.The collection includes the Intelligent Robotics Laboratory, Engineering Center, Boulder Creek Path, Edgar Experimental Mine, and ASPEN Lab.Source: Data Description · p. 4
Binary File Format
- Radar and lidar measurements use language-independent little-endian binary arrays, with Python and MATLAB parsing examples provided.Each binary file stores an array of numeric values; the organization varies by measurement type.Source: Binary File Format · p. 4
Raw Radar Data
- Raw radar frames store 16-bit signed integers, with sample, chirp, receiver, and transmitter indices mapped to array offsets and adjacent values representing I and Q.Python and MATLAB examples demonstrate parsing.Source: Raw Radar Data · p. 4
Radar Heatmaps
- Radar heatmaps store intensity and range rate for each range, azimuth, and elevation bin, with the Doppler dimension collapsed.The files contain 32-bit floating-point values; negative range rates indicate motion toward the sensor and positive rates motion away.Source: Radar Heatmaps · p. 5
Radar Pointclouds
- Radar pointclouds contain local-frame x, y, and z coordinates, intensity, and range rate; point counts vary by scan and are generally below 300.Each point comprises five contiguous 32-bit floating-point values.Source: Radar Pointclouds · p. 5
Lidar Pointclouds
- Each lidar point comprises local-frame x, y, and z coordinates and intensity, with 65,536 points per scan; IMU records contain linear acceleration and angular rates.Lidar points are stored as contiguous 32-bit floating-point values; IMU measurements are six space-delimited floating-point values per line.Source: Lidar Pointclouds · p. 5
Groundtruth
- Groundtruth poses come from a globally optimized pose graph constrained by IMU, lidar, and loop closures, and are expressed relative to each sequence’s starting pose.Each pose contains position and an orientation quaternion; the method is described in Hess et al. (2016).Source: Groundtruth · p. 5
Vicon
- Vicon poses are expressed relative to the Vicon coordinate frame and are available only for ASPEN Lab sequences.Each pose contains position and an orientation quaternion.Source: Vicon · pp. 5–6
Development Tools
Development tools provide loaders for sensor measurements and calibrations, plus an example script for visualizing readings and groundtruth poses.
Development Tools
- The development tools include example Python tools for reading and plotting dataset files, with MATLAB equivalents also provided.The tools are described in the supplied README.md file.Source: Development Tools · p. 6
- dataset loaders.py parses sensor measurements and calibrations into logical data structures, including heatmap arrays and cascade-radar parameter dictionaries.get heatmap() returns intensity and range-rate values in a NumPy array of dimension Ne × Na × Nr × 2.Source: Development Tools · p. 6
- plot pointclouds.py demonstrates an animation of sensor readings and groundtruth poses in a sensor-centric frame.The script uses functions from dataset loaders.py.Source: Development Tools · p. 6
Radar Calibration
Radar calibration covers antenna locations, waveform settings, antenna coupling, and phase and frequency corrections for the cascaded sensor. The paper also provides calibration files and code for applying several corrections.
Radar Calibration
- The provided antenna-location calibration maps each radar board’s TX and RX antenna positions, while waveform parameters include start frequency, frequency slope, and ADC sample rate.Configuration files are supplied for both cascade and single-chip sensors.Source: Radar Calibration · p. 6
- Antenna coupling calibration compensates for unintended voltage induced in adjacent antenna elements by correcting range-FFT outputs with calibration values recorded when no targets are present.The paper provides coupling calibration files and Python code for applying them.Source: Radar Calibration · p. 6
- The cascaded sensor receives phase and frequency calibration because its four chips may have small frequency, phase, and amplitude mismatches.The paper provides phase and frequency calibration vectors and Python code showing how to apply calibration.Source: Radar Calibration · pp. 6–7
- Frequency calibration uses a calibration vector to calculate a matrix, while the raw ADC samples for each chirp are frequency- and phase-calibrated before further processing.The calibration descriptions define sampling rate, frequency slope, and range-index values used in the process.Source: Radar Calibration · pp. 6–7
Extrinsic Calibration
Extrinsic transforms are specified relative to the IMU-coincident base frame, whose axes align with ENU. Sensor transforms are hand-measured, with lidar-to-radar alignment further refined using matched obstacle locations.
Extrinsic Calibration
- The base frame coincides with the IMU in translation and aligns with ENU, with x right, y forward, and z up.Transform files give translation followed by a quaternion in x, y, z, w order.Source: Extrinsic Calibration · p. 7
- Sensor extrinsics are hand-measured, and lidar-to-radar alignment is refined by matching two obstacles across lidar point clouds and radar heatmaps.The transforms do not change appreciably between runs, so one set is provided.Source: Extrinsic Calibration · p. 7
Summary and Future Work
ColoRadar provides diverse environments, 6DoF motion, and multiple formats of 4D radar measurements for robotic perception. The authors identify semantic labeling of lidar point clouds as a possible future extension.
Summary and Future Work
- The dataset includes indoor, outdoor, and subterranean environments, 6DoF motion, and three formats of 4D radar measurements from two sensors for odometry, mapping, and SLAM research.The authors state that its diversity makes it useful for autonomous robotics tasks ranging from disaster response to subterranean exploration.Source: Summary and Future Work · p. 7
- Adding semantic labels to lidar point clouds is identified as a potential way to extend the dataset toward object classification and semantic segmentation.The paper presents this as a future-work direction.Source: Summary and Future Work · p. 7