SLAM

How each robot builds a map of a building and keeps track of where it is on that map.

M20 Pro

The M20 has two ways to map and localize: DEEP Robotics’ 2D mapping stack, which is the one connected to navigation, and Lightning-LM, which maps and localizes in 3D. Lightning-LM runs beside the 2D stack on AOS, produces its own map and pose, and doesn’t drive the planner.

2D mapping

Provided by DEEP Robotics: a FAST-LIO-style filter for odometry and mapping, and GTSAM plus ICP for localization, fed by the RoboSense LiDAR (/LIDAR/POINTS) and Yesense IMU (/IMU_YESENSE).

• Separate estimates: localization reads raw LiDAR and IMU, not /LIO_ODOM, so two position estimates run at once.

2D mapping commands

Maps are stored in /var/opt/robot/data/maps, and active points to the one in use.

Status topics: /LOCATION_STATUS, /ODOM, /ALIGNED_POINTS, plus /SLAM_ODOM and /SLAM_ALIGNED_POINTS while mapping.

3D mapping: Lightning-LM

Lightning-LM (DEEP Robotics’ fork of gaoxiang12/lightning-lm, BSD-3) is another way to map and localize on the M20, this time in 3D. It is built on AOS at /home/user/lightning-lm.

• LiDAR-inertial odometry: a FAST-LIO-style filter fuses the IMU and LiDAR and tracks full 3D motion, with no flat-floor assumption.
• Scan skipping: LiDAR corrections run on every 2nd scan to save CPU, while the IMU still handles every scan.
• Loop closure: revisited places are matched and a pose graph removes accumulated drift.
• Start-anywhere localization: a place-recognition index lets the robot localize from anywhere in a saved map.
• Tiled maps: maps are saved as tiles, and distant tiles are unloaded during localization to save memory.

A Lightning-LM map (global.pcd) of the Levine hallway, viewed in pcl_viewer.

Installing. To install Lightning-LM on the M20, follow the steps in the DEEP Robotics Lightning-LM repository.

Before each session:

1. Start the point-cloud relay on NOS. It exists since V1.2.x but isn’t enabled at boot:

   ssh -J user@10.21.41.1 user@10.21.31.106
   sudo systemctl start multicast-relay.service
   

2. Open the AOS root shell. onboard.py refuses to run as user on the M20:

   ssh user@10.21.41.1
   cd ~/lightning-lm
   sudo bash
   export LIGHTNING_DIR=/home/user/lightning-lm && cd "$LIGHTNING_DIR"
   

Mapping, saving and localizing:

tmux new-session -s lightning -c "$LIGHTNING_DIR"
python3 scripts/onboard.py status                     # must print SUCCESS
python3 scripts/onboard.py slam                       # map; add --record-bag to keep its inputs
python3 scripts/onboard.py save run1                  # from a 2nd root shell; wait for "response: 0"
python3 scripts/onboard.py localize data/run1 --global-init

• Keep the robot still until it initializes, then walk it around. Revisiting earlier spots gives loop closure a chance.
• A map is saved only by save; stopping without it loses the map. Use a new name each run, because a non-empty folder won’t be overwritten.
• --global-init uses places.bin to start anywhere in the map, confirming over three observations. Without it, start near where mapping began.
• Stop with Ctrl+C and wait for the prompt. Localization writes accepted poses to data/localization_onboard.tum, replaced each run.

Map files

Topics

Viewing. Copy the map to a laptop and open it with pcl_viewer:

# on your laptop
mkdir -p ~/m20_runs
rsync -aP user@10.21.41.1:/home/user/lightning-lm/data/run1/ ~/m20_runs/map_run1/
pcl_viewer ~/m20_runs/map_run1/global.pcd -ps 2

Files are written by root, so run sudo chown -R user:user /home/user/lightning-lm/data on AOS first if copying fails. To visualize live in RViz, follow the steps in the DEEP Robotics Lightning-LM repository.

Troubleshooting

X30 Pro