Rover

Radar chirp data collection platform based on the TI AWR1843/DCA1000EVM.

This repository contains three components, each of which should run on a different machine:

• lidar: LIDAR/IMU data collection infrastructure for an Ouster LIDAR + Xsens IMU.
• radar: Radar data collection infrastructure for the TI AWR1843Boost + DCA1000EVM.
• processing: Radar/LIDAR/IMU data processing to create the final dataset.

Physical Hardware

• 1 Windows computer (GUI required) for Radar data collection.
• 1 Ubuntu 20.04 (focal) computer for Lidar/IMU data collection.
• AWR1843Boost Evaluation Board
  • 1 micro USB cable connecting the AWR1843's USB port to the windows computer.
  • 5v 3A power supply
  • NOTE: sketchy USB cables may cause the radar/capture card to fail to be detected.
• DCA1000EVM Capture Card
  • Powered via the AWR1843Boost.
  • 1 micro USB cable connecting the RADAR_FTDI port to the windows computer.
• Ouster OS0-64 LIDAR
  • 1 Cat5 ethernet cable connecting the LIDAR interface box to the linux computer.
• Xsens MTi-3 AHRS Development Kit
  • 1 micro USB cable connecting the IMU's USB port to the linux computer.
• 1 AC Battery Bank for powering the setup, with an additional power strip if required.

Control System Options:

1. RDP + SSH (Recommended for handheld operation):
   • 1 Windows laptop for controlling the Windows computer (via Windows RDP).
   • 1 (Wireless) Router that both computers and the laptop should connect to. Can be replaced with a wired solution if multiple LAN ports are available on each computer (adapters are ok).
   • Make sure to connect both computers to the same network, and assign them known host names (e.g. dart-lidar and dart-radar).
   • Set static IPs to the radar and lidar computers.
2. Manual Control (Recommended with a cart):
   • Use a laptop for the Windows and Linux computers, or connect external displays and keyboards to each.

Setup

See the setup guide.

Data Collection Guidelines

1. Keep a reasonable speed. DART will reject data when the speed exceeds 0.89 m/s (due to doppler "wrapping") or is less than 0.2 m/s (due to overly wide doppler bins). High speeds may lead to inconsistent velocity estimates by cartographer, so we suggest targeting 0.5m/s.
2. Move smoothly, again to assist cartographer in velocity estimation.
3. Keep the radar above your head when moving through tight spaces. The LIDAR has a minimum distance of 0.3m, and can lose tracking in tight spaces if you obscure half of its viewing range.
4. Scan high and low. The actual FOV of the radar is relatively narrow, so you should explicitly scan low and high points.
5. Emphasize occlusion. Basically any radar processing algorithm will work in a featureless metal box. Try to capture complex geometry and occlusion patterns (e.g. going behind walls, furniture, etc).

Usage

Note that these steps should be performed simultaneously on the Linux and Windows computer. In particular, make start and python collect.py should be performed right before the actual data collection step to avoid excess file size.

For a detailed step-by-step breakdown which bypasses any high-level automation for troubleshooting/development, see the manual data collection instructions.

Linux Computer, in the rover/lidar/ directory (collects 3.5GB/minute):

• On reboot (~15 seconds):
  1. Synchronize time with the windows computer: sudo ntpdate dart-radar.local
  2. Initialize ROS nodes: make init
• On each data collection (~30 seconds):
  1. Plug in the LIDAR. Wait until you can hear/feel the LIDAR reaching a steady state after spinning up.
  2. Start data collection: OUT=lidar.bag make start (replace lidar.bag with the desired output file name).
     • NOTE: it may take up to 30 seconds for data to start being collected; one way to check is to watch the file size of lidar.active.bag, and wait until it starts to increase rapidly.
     • Starting LIDAR data collection before starting radar data collection is suggested.
  3. Stop data collection: make stop. A large lidar.bag file (several GBs) should be created.
  4. Unplug the LIDAR.
• Cleanup: make deinit

Windows Computer, in the rover/radar directory (collects 1GB/minute):

• On reboot (~2 minutes):
  1. Power on the Radar, and make sure the XDS110 Class Application/User UART COM port matches what you have in config.json.
  2. Launch mmWave studio, and wait for all initialization commands to complete: python init.py
• On each data collection (~5 seconds):
  1. Run python collect.py. A radarpackets.h5 file should be created.
  2. Press ctrl+C once on python collect.py when finished. Do not close mmWave Studio, or you will need to restart the whole procedure and reflash the radar and capture card.
• Cleanup: close mmWave studio. The radar can stay powered on.

Data Processing, on a separate computer with cartographer-ros installed:

1. Copy the collected lidar and radar data to a folder; name the lidar data lidar.bag, and the radar data radarpackets.h5.

2. Run cartographer (~30 minutes):

   DIR=<dataset_directory> make lidar

   • Replace <dataset_directory> with the folder containing lidar.bag; all output files are also placed in this folder.
   • When running the makefile, the first step (roslaunch slam offline_cart_3d.launch ...) will wait indefinitely after it finishes if rviz is enabled. If this happens, close the rviz window, and the script will continue. DO NOT use ctrl+c; this will cancel the makefile as well.
   • Instead of passing DIR=..., you can alternatively copy this makefile to the <dataset_directory> and simply make.
   • This should create a number of files, including trajectory.csv and lidar.bag_points.ply.

3. Run radar processing & dataset packaging:

   DIR=<dataset_directory> make process

   • Replace <dataset_directory> with the folder containing trajectory.csv, lidar.bag_points.ply, and radarpackets.h5 from above (this step can be performed on a different machine, preferrably GPU-accelerated, as long as the above files are copied over).
   • This should create a number of files, including radar.h5, trajectory.h5, map.npz, and speed_report.pdf.