Collaborative SLAM Workshop#

This tutorial runs the Collaborative SLAM multi-camera tutorial in a workshop. Collaborative SLAM is a visual SLAM system in which several trackers, one per camera or robot, build maps and share them through a server. The workshop runs two trackers, one for each camera of a recorded two-camera bag, with ORB feature extraction on the Intel® GPU through Level Zero (SYCL). A UKF (robot_localization) fuses both trackers into one robot pose.

Prerequisites#

  • A host set up as described in Getting Started.

  • About 20 GiB free in Workshop’s storage pool. The GPU ORB extractor pulls in the full Intel® oneAPI toolkit, about 10 GB on disk.

  • A local desktop session for the RViz step.

Launch the workshop#

The launch takes about 13 minutes.

cd src/components/collaborative-slam/workshop
workshop launch collab-slam

Check the GPU#

workshop run collab-slam check-gpu

The Intel® Arc™ GPU must appear as a level_zero:gpu device in the sycl-ls output. If it doesn’t, the trackers can’t use the GPU.

Run the trackers without a display#

workshop run collab-slam tracker-check

The action replays the bag through both trackers and the UKF. It prints PASS if:

  • Both trackers publish kf_pose.

  • The UKF publishes /odometry/filtered.

  • Both tracker processes used the GPU, measured as compute-engine cycles from the xe driver.

  • Nothing crashed.

On Jazzy, the trackers publish 46 and 51 kf_pose messages and the UKF 55 fused poses in a 14-second replay.

View the trackers in RViz#

workshop connect collab-slam/robotics-ai-suite:desktop
workshop run collab-slam tutorial

The action runs the packaged cslam-multi-camera.sh: the UKF, static transforms, and two GPU trackers, each with its own RViz window, and plays the multi-camera bag (about 13 seconds). It runs until you press Ctrl+C.

Actions#

Action

Description

check-gpu

/dev/dri and sycl-ls. The GPU must appear as a level_zero:gpu device

tutorial

The packaged cslam-multi-camera.sh with RViz for each tracker. Runs until Ctrl+C

tracker-check

Headless: the same pipeline without RViz. PASS if both trackers publish, the UKF fuses them, both trackers used the GPU, and nothing crashed

Both pipeline actions add oneAPI’s umf/<version>/lib directory to LD_LIBRARY_PATH. Without it, the Level Zero adapter can’t load libumf.so.1, and the trackers fail with No device of requested type available.

Humble#

Use collab-slam-humble in every command. tracker-check passes: the trackers publish 46 and 48 kf_pose messages and the UKF 56 fused poses in a 12-second replay.

Ubuntu 22.04 has an older GPU compute runtime, and sycl-ls names the GPU Intel(R) Graphics [0xb080], but Level Zero and the GPU tracker work.

Known issues#

  • The -lze metapackage installs the CPU tracker. The generic ros-<distro>-univloc-tracker, -server and -slam packages (2.3-3 and 2.3-4) declare that they provide the -lze package names. So ros-<distro>-collab-slam-lze is satisfied by the CPU builds, and the tracker never loads the GPU ORB extractor. The workshop’s package list names the -lze packages explicitly, which replaces the generic ones. With the CPU build installed, the trackers still produce poses, but tracker-check reports FAIL because the GPU isn’t used.

  • Large install. liborb-lze depends on the full intel-oneapi-toolkit (81 packages, a 2.4 GB download and 10.2 GB on disk), although the oneAPI runtime packages would be enough.

  • UKF messages. The tutorial’s UKF node logs Could not obtain transform from odom->base_link a few times during the replay, and a NaN error after the bag ends. These come from the tutorial configuration and don’t affect the tracker output.

Source#

The definitions, package list and check script are in src/components/collaborative-slam/workshop/.